
VLSI Design
Size an inverter and watch its switching threshold move, race carries through fast adders, and find the vector that catches a stuck-at fault
24 chapters, 24 interactives, few words. Every chapter is mapped to the Anna University, VTU, SPPU, Mumbai University and KTU units it covers: MOSFET I-V and threshold, scaling, layout and lambda rules, the CMOS inverter's transfer curve, delay and power, static, pass-transistor and dynamic logic, wires and Elmore delay, logical effort, setup and hold timing, pipelining, clock distribution, metastability, fast adders, multipliers, SRAM, DRAM and flash, FPGAs, Verilog RTL design, the ASIC flow and yield, and testing.
- AUEC3552VLSI and Chip Design
- VTUBEC602VLSI Design and Testing
- SPPU404182VLSI Design and Technology
- MUECC503Digital VLSI
- KTUPEECT415Digital Systems and VLSI Design
By Skill-Lync · Updated 19 Nov 2026 · 24 chapters · 24 interactives · free, no sign-up
The MOS transistor: regions of operation and the I-V characteristics
Which syllabus covers thisAU EC3552 Unit IVTU BEC602 Module 2SPPU 404182 Unit IVMU ECC503 Module 1KTU not in syllabus
What are the three regions of operation of a MOSFET?
Below the threshold voltage VT (0.5 V in this book's illustrative process) there is no channel, so the transistor is off: cutoff. Above VT the gate pulls a thin layer of electrons to the silicon surface, joining source to drain. With a small VDS that layer behaves like a resistor whose value the gate sets: the triode (linear) region. Once VDS reaches the overdrive VGS − VT, the channel pinches off at the drain end and the current almost stops growing: saturation. With VGS = 1.3 V the overdrive is 0.8 V, so saturation starts at VDS = 0.8 V. A PMOS works the same way with every voltage and current reversed.
How large is the drain current, and why does it follow a square law?
The channel charge per unit area is Cox(VGS − VT − V), where V is the local channel voltage, and the electrons drift at a speed proportional to the field. Adding this up along the channel gives the triode equation below. It peaks where VDS equals the overdrive and stays there in saturation, at (k'/2)(W/L)(VGS − VT)². With k'n = 200 µA/V², W/L = 2 and VGS = 1.3 V, that is 128 µA. Doubling the overdrive gives four times the current, and doubling W gives twice. Holes are slower (k'p = 80 µA/V²), so a PMOS needs 2.5 times the width to match.
Why does the saturation current still creep up, and what are gm and rds?
As VDS rises, the pinch-off point moves towards the source and the effective channel gets shorter, so the current grows slightly. This is channel-length modulation, modelled as a factor (1 + λVDS). With λ = 0.1 per volt, moving VDS from 1.0 V to 1.8 V raises ID by 7.3 %. The transconductance gm = ∂ID/∂VGS tells you how strongly the gate controls the current: 320 µA/V at W/L = 2 and VGS = 1.3 V. The output resistance rds = 1/(λID) is 78.1 kΩ there, and their product, the most voltage gain one transistor can give, is 25.
What changes when the channel is short: velocity saturation?
Electron speed rises with the electric field only up to a limit. At L = 0.18 µm, 1.8 V across the channel is an average field of 10 V/µm, so carriers near the drain hit that limit. This book models it by dividing the triode current by 1 + VDS/(Ec·L), with an illustrative Ec·L of 1.08 V for electrons. At VGS = 1.8 V and W/L = 2, saturation now starts at 0.91 V instead of 1.3 V, and the current falls by 51 %. Holes saturate at higher fields, so velocity saturation also shrinks the nMOS/pMOS current ratio from 2.5 to 1.78.
An nMOS has VGS = 1.0 V (VT = 0.5 V) and λ = 0.1 per volt. You raise VDS from 1.0 V to 1.8 V. What happens to ID?
With an overdrive of 0.5 V the transistor is already saturated at VDS = 1.0 V, so the current is set by the gate, not the drain. Only channel-length modulation adds anything: (1 + 0.1 × 1.8)/(1 + 0.1 × 1.0) = 1.18/1.10, a 7.3 % rise.
Worked examples
Region and current at two drain voltages · nMOS, W/L = 3, VGS = 1.2 V, VT = 0.5 V, k'n = 200 µA/V²; VDS = 0.4 V and 1.5 V; first λ = 0, then λ = 0.1 per volt
- β = k'n·W/L = 600 µA/V², overdrive VGS − VT = 0.7 V, so saturation begins at VDS = 0.7 V
- VDS = 0.4 V < 0.7 V: triode. ID = 600 × (0.7 × 0.4 − 0.4²/2) = 600 × 0.20 = 120 µA
- VDS = 1.5 V > 0.7 V: saturation. ID = (600/2) × 0.7² = 147 µA
- With λ = 0.1 per volt multiply by (1 + λVDS): 120 × 1.04 = 124.8 µA and 147 × 1.15 = 169.05 µA
Answer: Triode, 120 µA (124.8 µA with λ) at 0.4 V; saturation, 147 µA (169 µA with λ) at 1.5 V
Transconductance, output resistance and intrinsic gain · nMOS, W/L = 4, VGS = 1.0 V, VDS = 1.2 V, λ = 0.1 per volt, square law
- β = 800 µA/V², overdrive 0.5 V < VDS: saturation
- I0 = (800/2) × 0.5² = 100 µA; with CLM ID = 100 × (1 + 0.12) = 112 µA
- gm = β(VGS − VT)(1 + λVDS) = 800 × 0.5 × 1.12 = 448 µA/V (a numeric derivative gives the same)
- rds = 1/(∂ID/∂VDS) = 1/(λ I0) = 1/(0.1 × 100 µA) = 100 kΩ
- Intrinsic gain gm·rds = 448 µA/V × 100 kΩ = 44.8
Answer: ID = 112 µA, gm = 448 µA/V, rds = 100 kΩ, gain ≈ 44.8
How much current does velocity saturation take away? · W/L = 2, VGS = 1.8 V, VDS = 1.8 V, λ left out; velocity-saturated model with Ec·L = 1.08 V (nMOS) and 2.7 V (pMOS), illustrative
- Long channel: VDsat = 1.3 V and ID = (400/2) × 1.3² = 338 µA
- Velocity saturated: the triode curve 400 × (1.3V − V²/2)/(1 + V/1.08) peaks at VDsat = 1.08 × (√(1 + 2 × 1.3/1.08) − 1) = 0.914 V (a golden-section search finds the same peak)
- ID = 166.9 µA, 50.6 % less than the square law predicts
- pMOS at the same W/L: 135.2 µA long-channel, 93.8 µA velocity saturated
- nMOS/pMOS current ratio: 2.5 without and 1.78 with velocity saturation
Answer: VDsat falls from 1.3 V to 0.91 V and ID from 338 to 167 µA; the n/p ratio drops to 1.78
Threshold voltage, body effect and MOS capacitances
Which syllabus covers thisAU EC3552 Unit IVTU BEC602 Module 2SPPU 404182 Unit IVMU ECC503 Module 1KTU not in syllabus
What sets the threshold voltage?
To form a channel in p-type silicon the gate must first cancel the built-in offset (the flat-band voltage VFB), then bend the bands by 2φF so the surface becomes as strongly n-type as the bulk is p-type, and finally hold up the charge of the depleted layer underneath, which adds γ√(2φF). So VT0 = VFB + 2φF + γ√(2φF). With this book's illustrative 2φF = 0.8 V and γ = 0.4 V^½, the last two terms add up to 1.158 V, so VT0 = 0.5 V needs VFB = −0.658 V. In a real process a threshold-adjust implant trims VT to the target.
What is the body effect?
If the source sits above the body (VSB > 0), the depletion layer under the gate is wider and holds more charge, so the gate needs a higher voltage to invert the surface. The threshold rises as in the formula below. With VT0 = 0.5 V, γ = 0.4 V^½ and 2φF = 0.8 V, VSB = 0.9 V gives VT = 0.664 V and VSB = 1.8 V gives 0.787 V. It matters wherever a source is not at ground: the upper nMOS in a NAND stack, and an nMOS pass transistor, whose output stops one raised threshold below its gate.
Where do γ and 2φF come from?
Both come from the substrate doping NA. The body-effect coefficient is γ = √(2qεSi·NA)/Cox: heavier doping puts more charge in the depletion layer, and a thinner oxide (larger Cox) lets the gate control that charge more easily. The surface potential is 2φF = 2(kT/q)·ln(NA/ni). For NA = 3 × 10¹⁷ cm⁻³ and tox = 4 nm (illustrative), γ = 0.366 V^½ and 2φF = 0.89 V. Thinner oxides in newer processes push γ down, so the body effect weakens as technology scales.
How does the gate capacitance split between source, drain and body?
The oxide capacitance is Cox = εox/tox = 8.63 fF/µm² for 4 nm of oxide, so a unit nMOS (W = 0.36 µm, L = 0.18 µm) has Cox·W·L = 0.559 fF. In cutoff the gate sees only the body (the full value in accumulation, less in depletion). In the linear region the channel connects to both ends, so the capacitance splits half to the source and half to the drain. In saturation the channel is pinched off at the drain, so about two thirds goes to the source and almost none to the drain. Overlap and junction capacitances come on top of these.
An nMOS has VT0 = 0.5 V, γ = 0.4 V^½ and 2φF = 0.8 V. Its source rises to 1.2 V above the body. What is the threshold now?
VT = 0.5 + 0.4 × (√2.0 − √0.8) = 0.5 + 0.4 × (1.414 − 0.894) = 0.708 V. The threshold grows with the square root of 2φF + VSB, so 1.2 V of source-body voltage raises it by only about 0.2 V.
Worked examples
Threshold of the upper transistor in a stack · nMOS with VT0 = 0.45 V, γ = 0.45 V^½, 2φF = 0.85 V; its source sits at VSB = 0.6 V, later 1.2 V
- √2φF = √0.85 = 0.922 V^½
- VSB = 0.6 V: √(0.85 + 0.6) = 1.2042, VT = 0.45 + 0.45 × (1.2042 − 0.922) = 0.577 V
- VSB = 1.2 V: √2.05 = 1.4318, VT = 0.45 + 0.45 × (1.4318 − 0.922) = 0.679 V
- Check: solving VGB = VFB + ψs + γ√ψs for the gate voltage that makes ψs = 2φF + VSB (bisection) gives the same values
Answer: VT = 0.577 V at VSB = 0.6 V and 0.679 V at 1.2 V (up 127 mV and 229 mV)
γ and 2φF from the doping and the oxide · p-substrate NA = 2 × 10¹⁷ cm⁻³, tox = 3 nm, ni = 10¹⁰ cm⁻³, kT/q = 25.85 mV
- Cox = 3.45 × 10⁻¹¹ / 3 × 10⁻⁹ = 11.51 fF/µm² (11.5104 × 10⁻³ F/m²)
- γ = √(2 × 1.602 × 10⁻¹⁹ × 1.036 × 10⁻¹⁰ × 2 × 10²³)/Cox = 0.224 V^½
- 2φF = 2 × 0.02585 × ln(2 × 10¹⁷/10¹⁰) = 0.869 V
- Body effect at VSB = 1 V: ΔVT = 0.224 × (√1.869 − √0.869) = 97 mV
- Check: integrating the depletion-layer field and bisecting its width gives the same charge per √ψ
Answer: γ = 0.224 V^½, 2φF = 0.869 V, ΔVT = 97 mV at VSB = 1 V
Gate capacitance of a wider transistor in each region · W = 1.2 µm, L = 0.18 µm, tox = 4 nm; intrinsic capacitances only
- Cox·W·L = 8.633 fF/µm² × 1.2 × 0.18 µm² = 1.865 fF
- Cutoff (accumulation): Cgb ≈ 1.865 fF, Cgs = Cgd ≈ 0
- Linear: Cgs = Cgd = 0.932 fF
- Saturation: Cgs = (2/3) × 1.865 = 1.243 fF, Cgd ≈ 0
- Check: differentiating the channel charge (integrated along the channel) gives Cgs = 0.667 CoxWL in saturation and 0.5 + 0.5 near VDS = 0
Answer: 1.86 fF to the body in cutoff; 0.93 + 0.93 fF when linear; 1.24 fF to the source in saturation
Technology scaling, short-channel effects and FinFETs
Which syllabus covers thisAU EC3552 Unit IVTU BEC602 Module 3SPPU 404182 Unit IVMU ECC503 Module 1KTU not in syllabus
What is constant-field (Dennard) scaling?
In 1974 Robert Dennard and colleagues showed that if every dimension and every voltage of a MOSFET shrinks by the same factor 1/S, the electric fields inside stay the same and the device behaves the same, only smaller and faster. One classic generation shrinks lengths to 0.7×, so S ≈ 1.43: each transistor takes 0.49 of the area, switches in 0.7 of the time and burns 0.49 of the power. Twice the transistors fit in the same area, and the power per square millimetre does not change.
What goes wrong if the supply voltage is not scaled?
Constant-voltage scaling shrinks the dimensions by 1/S but keeps VDD. The fields grow S times, the current per transistor grows S times, and so does its power, while S² more transistors share each square millimetre. Power density rises as S³. At S = 2 a gate is 4 times faster, but the same die area now dissipates 8 times the power: an illustrative 20 W chip would need to shed 160 W. The stronger fields also stress the gate oxide and create hot carriers. General scaling writes the supply as ε/S: ε = 1 is Dennard, ε = S is constant voltage.
Why did Dennard scaling stop?
The threshold voltage cannot keep shrinking. Below VT the current does not vanish: it falls by ten times for every 59.5 mV of gate voltage at best at room temperature, and about 89 mV in a typical device (illustrative factor 1.5). Lowering VT by 0.3 V would multiply the off-state leakage by about 2,300. So VT, and with it VDD, stopped falling in step with the dimensions, ε grew above 1 and power density started to rise. Dennard scaling broke down around 2005 to 2007, and designers turned to multicore chips and idle (dark) silicon.
What are short-channel effects?
When the channel is only a little longer than the depletion layers around the source and drain, the gate stops being fully in charge. VT roll-off: the junctions supply part of the channel's depletion charge, so VT falls as L shrinks. DIBL (drain-induced barrier lowering): a high drain voltage lowers the barrier at the source, so VT also falls as VDS rises and leakage grows. Velocity saturation: at L = 0.18 µm and 1.8 V the average field is 10 V/µm, enough to cap the carrier speed, so current grows almost linearly with overdrive. Hot carriers: fast electrons get into the oxide and slowly shift VT over the chip's life.
What do FinFETs, gate-all-around transistors and CNTFETs change?
A planar gate controls the channel from the top only. A FinFET stands the channel up as a thin fin and wraps the gate over three sides, so the gate holds the channel much more firmly: less DIBL, a steeper turn-off and lower leakage. Intel announced its 22 nm Tri-Gate (FinFET) transistors in 2011. Width now comes in whole fins: with an illustrative fin 40 nm tall and 8 nm wide, each fin gives 2 × 40 + 8 = 88 nm of width, so 360 nm needs 5 fins. Gate-all-around transistors surround stacked nanosheets on all four sides; carbon-nanotube FETs use a nanotube as the channel and are still mostly research.
You shrink every dimension of a chip by half (S = 2) but keep the supply voltage. What happens to the power per square millimetre?
Each transistor carries S = 2 times the current at the same voltage, so it burns 2 times the power, and S² = 4 times as many transistors fit in each square millimetre: 2 × 4 = 8 = S³. Only scaling the voltage too (constant field) keeps power density constant.
Worked examples
One constant-field generation · Start: L = 0.18 µm, VDD = 1.8 V (illustrative). Shrink every length and voltage to 0.7×
- S = 1/0.7 = 1.429; L = 0.126 µm, VDD = 1.26 V, field unchanged
- Gate capacitance C = εWL/tox scales as 1/S = 0.7; current I ∝ (W/L)(V²/tox) scales as 1/S = 0.7
- Delay CV/I: 0.7 × 0.7/0.7 = 0.7, so the clock can run 1.43 times faster
- Power per gate VI = 0.7 × 0.7 = 0.49; transistors per mm²: S² = 2.04
- Power density: 0.49 × 2.04 = 1, unchanged (checked by rebuilding the device from scaled W, L, tox and V)
Answer: Delay × 0.7, power per gate × 0.49, density × 2.04, power density × 1
Constant-voltage scaling and the heat problem · S = 2 with VDD kept at 1.8 V; a chip that dissipates an illustrative 20 W on the same die area
- Field and current per transistor: × 2; capacitance × 1/2; delay CV/I × 0.25
- Power per transistor VI: × 2; transistors per area: × 4
- Power density: 2 × 4 = 8 = S³
- The same die area now dissipates 20 × 8 = 160 W
Answer: Gates are 4 times faster but the chip would dissipate 160 W: constant-voltage scaling hits a heat wall
General scaling when the supply stops falling · S = 2 (L from 0.18 to 0.09 µm), but VDD only drops from 1.8 V to 1.2 V
- VDD scales as ε/S: 1.2/1.8 = ε/2, so ε = 1.333
- Delay 1/(εS) = 1/(1.333 × 2) = 0.375
- Power per gate ε³/S² = 0.593; energy per switch CV² = ε²/S³ = 0.222
- Power density ε³ = 2.37: it rises even though each gate uses less energy
Answer: ε = 1.333: delay × 0.375, energy per switch × 0.222, power density × 2.37
CMOS fabrication, lambda design rules, stick diagrams and layout
Which syllabus covers thisAU EC3552 Units II, VVTU BEC602 Module 3SPPU 404182 Unit VMU ECC503 Modules 1-2KTU not in syllabus
How is a CMOS inverter built on the wafer?
Start with a lightly doped p-type wafer. Implant and drive in an n-well where the pMOS will sit. Define the active areas and isolate them with thick field oxide (shallow trenches in modern processes). Grow the thin gate oxide, deposit polysilicon and etch it into gates. Implant n+ and p+ for the sources and drains, using the gate itself as the mask, so the gate lines up with its own source and drain automatically (self-alignment). Cover everything in oxide, open contact holes and deposit and pattern metal. Our inverter layout uses 7 mask layers: n-well, active, poly, n-select, p-select, contact and metal1.
What are lambda design rules?
Design rules are the minimum widths, spacings and overlaps the masks can reliably reproduce. The MOSIS scalable CMOS (SCMOS) rules write every one as a multiple of a unit λ, so one layout can be mapped onto several processes. Poly is at least 2λ wide, metal1 3λ, a contact exactly 2λ × 2λ, and active must surround a contact by 1.5λ. MOSIS maps λ to 0.35 µm (SCMOS) or 0.30 µm (SUBM) on ON Semi's C5 process, and to 0.09 µm (DEEP) on TSMC 0.18 µm. SUBM and DEEP enlarge the rules that did not shrink as fast as the gate. Shapes must sit on a half-λ grid.
What is a stick diagram, and what does the inverter's look like?
A stick diagram shows which layers connect, and in what order, without any dimensions. The usual colour code is green for diffusion (active), red for polysilicon, blue for metal1, a black cross or dot for a contact, and a dashed outline for the n-well. The inverter: a blue VDD line across the top and a blue GND line across the bottom. Two green strips: the pMOS near VDD inside the well, the nMOS near GND. One red line runs down through both, and each crossing of red over green is a transistor. The left ends contact the rails, and a blue line joins the two right ends to form the output.
How big is the inverter, and what checks a layout?
Drawn to the SUBM rules at minimum, the figure's inverter is 25 λ wide and 45.5 λ tall. At λ = 0.30 µm that is 102.4 µm². The same drawing at the DEEP λ of 0.09 µm takes 9.21 µm², because area scales as λ². Three checks run before tape-out. Design-rule checking (DRC) measures every width, spacing and overlap. Layout-versus-schematic (LVS) extracts the transistors and wires from the layout and compares them with the schematic. Parasitic extraction adds the wire resistances and capacitances for timing.
You move a layout from λ = 0.30 µm to λ = 0.15 µm, keeping every rule the same in λ. What happens to the cell area?
Every width and height halves in microns, so the area falls by 2 × 2 = 4. The 25 λ × 45.5 λ inverter drops from 102.4 µm² to 25.6 µm². That is what makes λ rules scalable.
Worked examples
Turn λ rules into microns · SCMOS_SUBM rules at λ = 0.30 µm: poly width 2λ, contact 2λ × 2λ, contact to gate 2λ, active around a contact 1.5λ, metal1 width 3λ and spacing 3λ
- Gate length (poly width): 2 × 0.30 = 0.60 µm; contact: 0.60 µm square
- Metal1: 0.90 µm wide with 0.90 µm gaps, a 1.8 µm pitch for minimum wires
- Source-contact centre to drain-contact centre: 1 + 2 + 2 + 2 + 1 = 8 λ = 2.4 µm
- Active length of the contacted transistor: 1.5 + 2 + 2 + 2 + 2 + 2 + 1.5 = 13 λ = 3.9 µm
Answer: 0.6 µm gate, 0.9 µm metal, 2.4 µm between contacts, 3.9 µm of active per transistor
Find the violations in a proposed layout · The book's inverter in SCMOS_SUBM with poly width 1.5λ, gate extension 2λ, contact to gate 1.5λ, rail width 3λ and nMOS width 4λ
- 3.1 poly width: 1.5λ < 2λ, fails
- 6.2 active around a contact: the 4λ nMOS active leaves (4 − 2)/2 = 1λ above and below the 2λ contact, < 1.5λ, fails
- 6.4 contact to gate: 1.5λ < 2λ, fails
- 7.2 metal1 spacing: the 4λ-wide straps over the source and drain contacts sit 2 + 2 × 1.5 + 1.5 = 6.5λ apart, leaving 2.5λ of gap < 3λ, fails
- Passing: contact spacing 4.5λ ≥ 3λ, gate extension 2λ ≥ 2λ; a brute-force rectangle check finds exactly these four failures
- Fix: poly 2λ, contact to gate 2λ, nMOS width 5λ: all 14 rules pass and the cell grows from 88.8 to 102.4 µm²
Answer: Rules 3.1, 6.2, 6.4 and 7.2 fail; the fixed cell takes 102.4 µm²
The same inverter in three rule sets · Poly 2λ, gate extension 2.5λ, contact to gate 2λ, rails 3λ, nMOS 5λ, pMOS 12.5λ; λ = 0.35 µm (SCMOS), 0.30 µm (SUBM), 0.09 µm (DEEP)
- Width = active length 7 + 2 × 2 + 2 = 13λ plus the well margin (rule 2.3: 5λ SCMOS, 6λ SUBM and DEEP) on both sides
- SCMOS: 23 × 43.5 λ = 1000.5 λ² × 0.35² = 122.6 µm²
- SUBM: 25 × 45.5 λ = 1137.5 λ² × 0.30² = 102.4 µm²
- DEEP: 25 × 45.5 λ × 0.09² = 9.21 µm²
- The SUBM well rule costs 2λ in each direction, but its smaller λ still wins
Answer: 122.6 µm² (SCMOS), 102.4 µm² (SUBM), 9.21 µm² (DEEP)
The CMOS inverter: switching threshold, sizing and noise margins
Which syllabus covers thisAU EC3552 Unit IVTU BEC602 Module 2SPPU 404182 Unit IVMU ECC503 Module 2KTU not in syllabus
What do the two transistors do as the input rises from 0 to VDD?
Take this book's illustrative process: VDD = 1.8 V, VTn = 0.5 V, VTp = −0.5 V. Below 0.5 V the NMOS is off, so the PMOS holds the output at exactly 1.8 V. Once the NMOS turns on, it is saturated while the PMOS is in its linear region, and the output sags a little. Around the switching threshold both transistors are saturated, so a tiny input change swings the output a long way. Past it the roles swap: the NMOS goes linear, and above 1.3 V the PMOS is off and the output is 0 V. Both rails are reached exactly, with no static current at either end.
Where does the inverter switch, and how does sizing move it?
VM is the input at which Vout = Vin. There both transistors are saturated, so equating their currents gives the formula below. Here r is the square root of the PMOS-to-NMOS strength ratio. Electrons are 2.5 times as mobile as holes in this process, so Wp/Wn = 2.5 makes r = 1 and VM = 0.9 V, half of VDD. A minimum-size PMOS (Wp = Wn) gives r = 0.632 and VM = 0.81 V. Because of the square root, VM moves slowly: doubling the PMOS from 2.5 to 5 raises it by only 69 mV.
What are VIL, VIH and the noise margins?
VIL and VIH are the two inputs where the transfer curve has slope −1. Between them a disturbance on the input is amplified, so a logic input must stay below VIL to count as 0 and above VIH to count as 1. The noise margins are NML = VIL − VOL and NMH = VOH − VIH, with VOL = 0 and VOH = VDD for static CMOS. For the symmetric inverter with λ = 0, VIL = (3VDD + 2VT)/8 = 0.80 V and VIH = 1.00 V, so both margins are 0.80 V. With λ = 0.1 V⁻¹ each becomes 0.783 V.
Why is the gain finite, and what does velocity saturation change?
With an ideal square law, both saturated transistors act as perfect current sources, so the curve is vertical at VM. Channel-length modulation gives each a finite output resistance. For the symmetric inverter the gain at VM is −2(1 + λVM)/(λ(VM − VTn)) = −54.5. Velocity saturation cuts the NMOS current more than the PMOS current in our model, so less PMOS widening is needed: VM = VDD/2 at Wp/Wn = 1.75 instead of 2.5. The gain at VM falls to about −35, because the saturated current grows more slowly with gate voltage.
A symmetric inverter (Wp/Wn = 2.5, VM = 0.9 V at VDD = 1.8 V) gets its PMOS doubled to Wp/Wn = 5. Where does VM go?
VM depends on the square root of the strength ratio. Doubling Wp makes r = √2 = 1.414, not 2, and VM = (0.5 + 1.414 × 1.3)/2.414 = 0.969 V. Moving the threshold a long way takes very lopsided sizing.
Worked examples
Switching threshold of a minimum-size inverter · Wn = Wp = 0.36 µm, L = 0.18 µm, VDD = 1.8 V, k'n = 200 µA/V², k'p = 80 µA/V², VTn = 0.5 V, VTp = −0.5 V (λ = 0)
- r = √(k'p·Wp/(k'n·Wn)) = √(80/200) = 0.6325
- VM = (VTn + r(VDD − |VTp|))/(1 + r) = (0.5 + 0.6325 × 1.3)/1.6325 = 0.8099 V
- Check: at VM both transistors are saturated (VDS ≥ VGS − VT for each), so the formula applies
- With λ = 0.1 V⁻¹ (the interactive) bisection of the current balance gives 0.8115 V, almost the same
Answer: VM = 0.81 V, about 90 mV below VDD/2: the weak PMOS lets the NMOS win early
Sizing the PMOS for a chosen threshold · The same process; the inverter must switch at VM = 0.95 V with VDD = 1.8 V
- Solve the VM formula for r: r = (VM − VTn)/(VDD − |VTp| − VM) = 0.45/0.35 = 1.2857
- r² = k'p·Wp/(k'n·Wn), so Wp/Wn = r² × k'n/k'p = 1.6531 × 2.5 = 4.133
- With Wn = 0.36 µm: Wp = 1.488 µm
- In the velocity-saturated option the current goes as VGT^1.3, so r^1.3 replaces r²: Wp/Wn = 2.42
Answer: Wp/Wn = 4.13 (Wp = 1.49 µm), or 2.42 with velocity saturation
Noise margins of the minimum-size inverter · Wn = Wp (βn/βp = kR = 2.5), VDD = 1.8 V, VT = 0.5 V, λ = 0
- At VIL the NMOS is saturated and the PMOS linear. Differentiating the current balance and setting dVout/dVin = −1 gives VDD − Vout = (b − kR·a)/2, with a = Vin − VTn and b = VDD − Vin − |VTp|
- Substituting back gives (4kR + kR²)a² + 2kR·ab − 3b² = 0, so a/b = 0.3025. With a + b = 0.8 V: VIL = 0.6858 V (Vout = 1.725 V)
- At VIH the roles swap: 3kR·a² − 2ab − (4 + 1/kR)b² = 0, so a/b = 0.9108 and VIH = 0.8813 V (Vout = 0.107 V)
- NML = VIL − 0 = 0.686 V, NMH = 1.8 − VIH = 0.919 V. Bisection on a numerical slope gives the same points
Answer: NML = 0.69 V, NMH = 0.92 V: the low-side margin shrinks when the PMOS is weak
Inverter propagation delay, load capacitance and fan-out
Which syllabus covers thisAU EC3552 Unit IIVTU BEC602 Module 3SPPU 404182 Unit IVMU ECC503 Module 2KTU not in syllabus
How is propagation delay measured?
Apply an edge to the input and time how long the output takes to cross the 50 % level, VDD/2. When the output falls this is tpHL, set by the NMOS discharging the load. When it rises it is tpLH, set by the PMOS charging it. Their mean is the gate delay tp. With the illustrative process, a unit inverter (Wn = 0.36 µm, Wp = 0.9 µm) driving four copies of itself has tpHL = tpLH = 27.0 ps. They are equal because Wp/Wn = 2.5 exactly cancels the mobility ratio.
What makes up the load capacitance?
Three parts. First, the inverter's own drain junctions: about 1 fF per µm of width, or 1.26 fF for the unit inverter. Second, the gates it drives: 2 fF per µm, or 2.52 fF per unit inverter of fan-out. Third, the wire, at 0.2 fF per µm here. Fan-out of four (FO4) gives CL = 1.26 + 4 × 2.52 = 11.34 fF. For a step input the delay is exactly proportional to CL, so going from FO4 to FO8 multiplies it by 1.889.
How does the RC estimate work?
Treat the on transistor as a resistor and use the RC charging result: the output crosses 50 % after ln 2 ≈ 0.69 time constants. The transistor is not linear, so take its average VDS/ID over the half swing from VDD to VDD/2 as Reff. For the unit NMOS that is 3.55 kΩ. With the FO4 load, 0.69·Reff·CL = 27.7 ps, against 27.0 ps from simulation. That is close enough to reason with.
Does a bigger inverter or a higher supply always make it faster?
A wider inverter drives more current, but it also brings more drain capacitance of its own. Driving a fixed 50 fF load, a 4× wider unit inverter is 3.73× faster, not 4×. It can never get faster than its intrinsic delay with no external load, 3.0 ps, and it loads whatever drives it 4× more (chapter 10 balances this). Supply matters too: dropping VDD from 1.8 to 1.2 V makes the FO4 delay 2.36× longer, because the drive current falls with (VDD − VT)².
A unit inverter drives four copies of itself (FO4). The fan-out is doubled to eight. What happens to the delay?
The delay is proportional to the total load, and only the fan-out part doubles. CL goes from 1.26 + 4 × 2.52 = 11.34 fF to 1.26 + 8 × 2.52 = 21.42 fF, a ratio of 1.889. The inverter's own drain capacitance does not grow.
Worked examples
Closed-form fall delay of a unit inverter at FO4 · Unit inverter βn = k'n·Wn/L = 200 × 2 = 400 µA/V², VDD = 1.8 V, VTn = 0.5 V, λ = 0, step input, CL = 11.34 fF
- While Vout > VDD − VTn = 1.3 V the NMOS is saturated with I = ½βn(VDD − VTn)² = 338 µA, so the first 0.5 V takes t1 = CL × 0.5/I = 16.78 ps
- Below 1.3 V it is linear: CL dV/dt = −βn((VDD − VTn)V − V²/2). Integrating from 1.3 V to 0.9 V gives t2 = CL/(βn·VGT) × ln((4VGT − VDD)/VDD)
- t2 = 11.34 fF/(400 µA/V² × 1.3 V) × ln(3.4/1.8) = 13.87 ps
- tpHL = t1 + t2 = 30.64 ps (an RK4 simulation gives the same)
Answer: tpHL = 30.6 ps
Equivalent resistance and the 0.69 RC estimate · The same unit NMOS with λ = 0.1 V⁻¹, gate at 1.8 V, output swinging from 1.8 V to 0.9 V, CL = 11.34 fF
- Reff = (1/0.9 V) × ∫ V/I(V) dV from 0.9 to 1.8 V (Simpson) = 3545 Ω (with λ = 0 it would be 4038 Ω)
- 0.69 × Reff × CL = 0.69 × 3545 Ω × 11.34 fF = 27.74 ps
- Simulated tpHL with λ = 0.1 V⁻¹: 27.03 ps
- A cruder rule, Reff ≈ ¾VDD/Isat = 3385 Ω, gives 26.5 ps
Answer: Reff ≈ 3.55 kΩ; the RC estimate 27.7 ps is within 3 % of the simulated 27.0 ps
Fan-out of three plus a wire · A unit inverter drives three unit inverters through 100 µm of metal (0.2 fF/µm); λ = 0.1 V⁻¹, VDD = 1.8 V
- CL = Cself + 3 Cin + Cwire = 1.26 + 3 × 2.52 + 100 × 0.2 = 28.82 fF
- For a step input the delay is CL × (∫dV/I), so it scales with CL: 2.383 ps per fF for this driver
- tpHL = tpLH = 28.82 × 2.383 = 68.7 ps
- The wire alone (20 fF) is 69 % of the load
Answer: tp ≈ 69 ps, 2.54× the FO4 delay; the wire, not the fan-out, dominates
Power dissipation in CMOS and low-power design
Which syllabus covers thisAU EC3552 Units I-IIVTU BEC602 Module 3SPPU 404182 Unit IVMU ECC503 Module 2KTU not in syllabus
Where does most of the power in a CMOS chip go?
Into charging and discharging capacitance. Each time a node goes from 0 to 1, the supply delivers CL·VDD² of energy. If α is the fraction of clock cycles with such a transition, the switching power is the formula below. A gate with 20 fF of load at 1.8 V, clocked at 500 MHz with α = 0.1, uses 3.24 µW. Power goes as VDD squared, so running the same gate at 1.2 V cuts it to 1.44 µW, which is 44 % as much.
Where exactly does the energy CL·VDD² end up?
While the PMOS charges the load, the supply delivers charge CL·VDD at the full voltage VDD, which is CL·VDD² = 64.8 fJ for 20 fF at 1.8 V. Only half of it, ½CL·VDD² = 32.4 fJ, ends up stored on the capacitor. The other half heats the PMOS, whatever its size or resistance. A simulation gives 32.4 fJ of heat. On the next 1→0 transition the NMOS burns the stored half. A wider transistor charges faster, but it wastes the same energy.
What is short-circuit power and when does it matter?
During an input edge there is a moment when both transistors are on, and current flows straight from VDD to ground. For an unloaded symmetric inverter with input ramp time tr, the energy per edge is β·tr·(VDD − 2VT)³/24. That is 853 aJ for the unit inverter with a 100 ps edge. A real load slows the output, so less current overlaps: with 20 fF the simulation gives 116 aJ, only 0.18 % of CL·VDD². Slow input edges increase it. If VDD is below VTn + |VTp| = 1.0 V, the two transistors are never on together, and it vanishes.
Why does a chip use power even when nothing switches?
Transistors that are off still leak: subthreshold current through the channel, and gate and junction leakage. Static power is VDD × Ileak per gate, independent of clock and activity. Take 20 nA per gate: that is 36 nW at 1.8 V. Now compare a gate with 15 fF of load at 100 MHz: its dynamic power equals the leakage when α = 0.007. A block that is mostly idle is therefore dominated by leakage. That is why designs use power gating (cutting the supply to idle blocks) and high-threshold transistors off the critical paths.
How do designers trade speed for energy?
Lower VDD cuts energy per transition as VDD², but delay rises as the drive current falls. The product, energy × delay, has a minimum: for this book's unit inverter it is at VDD = 1.57 V. There the energy is 76 % of its 1.8 V value and the delay 1.3× longer. Other levers cut α or f instead of VDD. Clock gating stops the clock to idle registers. Dynamic voltage and frequency scaling lowers VDD and f together when the work allows.
A block runs at 1.8 V. The supply is lowered to 1.2 V at the same clock frequency, and it still meets timing. What fraction of the switching power is left?
Switching power goes as VDD²: (1.2/1.8)² = 0.444. The frequency, load and activity are unchanged. Lowering f too would cut it further, to 22 % at half the clock.
Worked examples
Switching power of a logic block · 20 000 gates, average load 15 fF each, activity α = 0.1, clock 400 MHz, VDD = 1.8 V (illustrative block)
- Per gate: αCL·VDD²·f = 0.1 × 15 fF × (1.8 V)² × 400 MHz = 1.944 µW
- Block: 20 000 × 1.944 µW = 38.88 mW
- At 1.2 V and 200 MHz (a slower mode): 8.64 mW, 22 % of the full-speed power
- Energy per clock cycle at full speed: 38.88 mW/400 MHz = 97.2 pJ
Answer: 38.9 mW at full speed, 8.6 mW in the slow mode
Short-circuit energy: unloaded upper bound and a real load · Unit inverter (β = 400 µA/V² for both transistors), VDD = 1.8 V, VT = 0.5 V, input ramp 200 ps (0 to 100 %); load FO4 = 11.34 fF
- Unloaded, the output follows the static curve. Current flows while Vin is between 0.5 V and 1.3 V, and it is the NMOS saturation current up to VM = 0.9 V, symmetric after it
- E = 2 × VDD × ∫ ½β(Vin − VT)² dt = β·tr·(VDD − 2VT)³/24 = 400 µA/V² × 200 ps × 0.8³/24 = 1.707 fJ
- With the FO4 load (RK4 simulation of the ramp, λ = 0.1 V⁻¹): 0.584 fJ
- Switching energy for comparison: CL·VDD² = 36.74 fJ, so the unloaded bound is 4.6 % and the loaded value 1.6 %
Answer: 0.58 fJ per edge with the load (1.71 fJ unloaded), a few % of CV² at most
When leakage beats switching · One gate: CL = 15 fF, VDD = 1.8 V, f = 100 MHz, leakage 20 nA (illustrative)
- Leakage power: VDD × Ileak = 1.8 V × 20 nA = 36 nW, every cycle, switching or not
- Switching power per unit activity: CL·VDD²·f = 15 fF × 3.24 V² × 100 MHz = 4860 nW
- Equal when α = 36/4860 = 0.0074
- At α = 0.01 (a mostly idle block) switching is 48.6 nW, so leakage is 43 % of the total
Answer: Leakage dominates below α ≈ 0.007; for idle blocks, power gating saves more than clock gating
Static CMOS gates: NAND, NOR, complex gates and transistor sizing
Which syllabus covers thisAU EC3552 Unit IIVTU BEC602 Module 4SPPU 404182 Unit IVMU ECC503 Module 2KTU not in syllabus
How is any static CMOS gate built?
Every static CMOS gate has two networks. The pull-down network of NMOS transistors connects the output to ground, and it conducts exactly when the output should be 0. The pull-up network of PMOS transistors connects it to VDD and is the dual: NMOS in series become PMOS in parallel, and the reverse. NMOS in series act as AND and in parallel as OR, so the pull-down realises NOT f. For every input vector exactly one network conducts. A gate with n inputs uses 2n transistors, and there is no static current.
How wide should each transistor be?
Size each gate to match the unit inverter's worst-case drive: NMOS 1 unit, PMOS 2.5 units, where 1 unit = 0.36 µm. Two transistors in series have twice the resistance, so each must be twice as wide. The rule is width = unit × the longest series chain the transistor sits in. A NAND2 has NMOS of 2 and PMOS of 2.5, 9 units in total. A NOR2 has NMOS of 1 and PMOS of 5, 12 units. If more paths conduct at once the gate is stronger than this, never weaker.
Why are NAND gates preferred over NOR gates?
Logical effort g compares a gate's input capacitance with that of an inverter of the same drive. For one input it is (NMOS width + PMOS width)/3.5. A NAND2 input drives 2 + 2.5 units, so g = 1.286. A NOR2 input drives 1 + 5, so g = 1.714. The NOR stacks the slower PMOS in series, and that gets worse with more inputs: NAND3 g = 1.571, NOR3 g = 2.429. Published tables quote 4/3 and 5/3 because they assume a PMOS twice the NMOS width; ours is 2.5×.
What are AOI and OAI gates, and how tall can a stack get?
An AND-OR-INVERT gate computes NOT(AB + CD) in a single stage. Its pull-down is two series pairs in parallel, and its pull-up is two parallel pairs in series. That takes 8 transistors, against 14 for NAND-NAND logic plus an inverter, and there is only one gate delay. Tall stacks hurt, though. Each extra series transistor widens all of them, adds capacitance at the internal nodes and suffers body effect, because its source sits above ground. A common rule of thumb keeps stacks to 3 or 4 and splits bigger functions into two stages.
NAND3 and NOR3 are both sized for the unit inverter's worst-case drive. Which needs more total transistor width?
NAND3 has three NMOS in series (3 units each) and three PMOS in parallel (2.5 each): 16.5 units. NOR3 puts its three PMOS in series, each 3 × 2.5 = 7.5 units wide, plus three NMOS of 1: 25.5 units. That is 1.55× as much.
Worked examples
Sizing a three-input NAND · Unit inverter: Wn = 0.36 µm, Wp = 0.9 µm (mobility ratio 2.5); NAND3 for the same worst-case drive
- Pull-down: A, B and C in series, so each NMOS is 3 units = 1.08 µm wide
- Pull-up: three PMOS in parallel; the worst case is one on alone, so each is 2.5 units = 0.9 µm
- Total width = 3 × 3 + 3 × 2.5 = 16.5 units = 5.94 µm
- Each input sees 3 + 2.5 = 5.5 units, against 3.5 for the unit inverter: g = 1.5714
Answer: NMOS 1.08 µm, PMOS 0.9 µm, g = 1.571
Sizing an AOI22 gate · Y = NOT(AB + CD), the same unit inverter
- Pull-down: (A series B) in parallel with (C series D). Every NMOS is in a chain of 2, so each is 2 units
- Pull-up (the dual): (A parallel B) in series with (C parallel D). Every path has 2 PMOS, so each is 2 × 2.5 = 5 units
- Total = 4 × 2 + 4 × 5 = 28 units = 10.08 µm of gate width
- Each input drives 2 + 5 = 7 units: g = 7/3.5 = 2. Check: with only A and B on, R = ½ + ½ = 1 unit
Answer: NMOS 2 units (0.72 µm), PMOS 5 units (1.8 µm), g = 2 per input, 10.08 µm in total
Pattern-dependent delay of a NAND2 · NAND2 sized as above, driving 15 fF; the unit PMOS (and each NAND2 PMOS, 0.9 µm) has Reff from chapter 6 (λ = 0.1 V⁻¹, VDD = 1.8 V)
- Reff of a 2.5-unit PMOS = Reff of the unit NMOS = 3545 Ω (same βp = βn)
- Only A falls (B stays 1): one PMOS pulls up, tpLH ≈ 0.69 × 3545 Ω × 15 fF = 36.7 ps
- A and B both fall: two PMOS in parallel, R halves, tpLH ≈ 18.3 ps
- The pull-down has only one case: both NMOS (2 units each) in series, R = ½ + ½ = 1 unit, the same as the inverter
Answer: tpLH ≈ 37 ps (one input switching) or 18 ps (both): sizing only guarantees the worst case
Interconnect: wire resistance and capacitance, wire models and Elmore delay
Which syllabus covers thisAU EC3552 Units II, IVVTU BEC602 Module 3SPPU 404182 Unit IIMU ECC503 Module 5KTU not in syllabus
How much resistance and capacitance does an on-chip wire have?
A wire's resistance is its sheet resistance times the number of squares, length divided by width. Its capacitance is roughly proportional to length, to the layers above and below and to neighbouring wires. This book uses illustrative values of r = 0.08 Ω/µm and c = 0.2 fF/µm. So 1 mm of wire has 80 Ω and 200 fF. That capacitance equals the inputs of about 79 unit inverters (2.52 fF each), so a long wire is a big load before its resistance matters at all.
When is a wire just a capacitor, and when do you need a distributed RC model?
If the wire's resistance is small next to the driver's, lump all of its capacitance at one node: delay ≈ 0.69 Rd(Cw + CL). If not, the charge has to flow through the wire's own resistance, which is spread along it. Model it as N sections of R/N and C/N. A π section (C/2, R, C/2) gives the exact Elmore delay for any N. An L section (R then C) overestimates it with few sections. For the open-ended line with an ideal driver the exact 50% delay is 0.3787 RC, while one lumped RC predicts 0.69 RC.
What is the Elmore delay of an RC ladder?
Walk from the driver to the far end. For every capacitor, multiply it by the resistance it shares with that path, and add the products. That is Elmore's 1948 estimate: the time constant of the single exponential with the same area above it as the real response. For a driver Rd, a wire Rw, Cw and a load CL it gives the formula below. A 5 mm wire driven through 1 kΩ into 20 fF has τ = 1.228 ns. The wire's own term rises with the square of length.
Why do repeaters make a long wire faster?
A wire's own delay grows as L², because both R and C grow with length. Cut the wire into m pieces, each driven by a fresh buffer, and each piece's RC term falls by m². You pay one buffer delay per piece, so total delay becomes nearly linear in length, with a best m. Take a 10 mm wire driven by 20× inverters (200 Ω, 50 fF). Unbroken, the 50% delay is 918 ps. Cut into 9 pieces with 8 repeaters, it is 433 ps.
A long wire is driven by a very strong driver into almost no load. You double its length. What happens to its delay?
Both the wire's resistance and its capacitance double, and its own delay goes as R·C, so it rises about 4×. For our metal, 5 mm of wire gives 151 ps and 10 mm gives 606 ps (exact distributed-line 50% delay). Repeaters are how you get back to linear growth.
Worked examples
Elmore delay and 50% delay of a driven wire · A 3 mm wire (r = 0.08 Ω/µm, c = 0.2 fF/µm), driver Rd = 2 kΩ, load CL = 20 fF
- Rw = 0.08 × 3000 = 240 Ω, Cw = 0.2 × 3000 = 600 fF
- Driver term Rd(Cw + CL) = 2000 × 620 fF = 1240 ps
- Wire term RwCw/2 = 72 ps, load term RwCL = 4.8 ps
- Elmore τ = 1316.8 ps
- 50% estimate: 0.69 × 1240 ps + 0.38 × RwCw + 0.69 × RwCL = 913.6 ps
- A 20-section π-ladder simulation (RK4) gives 920.6 ps; the driver dominates, so the 0.69 weighting is close
Answer: τ = 1.317 ns; 50% delay ≈ 921 ps (estimate 914 ps)
Lumped versus distributed: a wire with an ideal driver · A 5 mm wire, ideal driver (Rd = 0), no load
- Rw = 400 Ω, Cw = 1 pF, RwCw = 400 ps
- Treating it as one R and one C: 0.69 RC = 276 ps (too pessimistic)
- Elmore τ = RwCw/2 = 200 ps
- Exact series solution of the distributed RC line: the far end reaches 50% at 0.3787 RC = 151.5 ps; 0.38 RC is the usual rounding
- A 40-section π ladder simulated with RK4 gives 151.5 ps
- Double the length to 10 mm: RC rises 4×, so the delay becomes 606 ps
Answer: About 151 ps; one lumped RC would overestimate it by 82%
How many repeaters for a 10 mm wire? · A 10 mm wire; repeaters are 20× unit inverters (Reff = 3.99 kΩ/20, Cin = 20 × 2.52 fF); the far end drives one more repeater
- Repeater: Rb = 199.7 Ω, C0 = 50.4 fF; wire Rw = 800 Ω, Cw = 2 pF
- No repeaters: 0.69 Rb(Cw + C0) + 0.38 RwCw + 0.69 RwC0 = 918.4 ps
- With m pieces: t(m) = 0.69 RbCw + 0.38 RwCw/m + (m − 1)(0.69 RbC0 + 0.69 RwC0/m) + 0.69(Rb + Rw/m)C0
- Best real m = √(0.38 RwCw / (0.69 RbC0)) = 9.36; checking whole numbers, m = 9 is best
- t(9) = 433.5 ps with 8 repeaters
- For 20 mm: 3046 ps unbroken, 867 ps with 18 repeaters, about twice the 10 mm value
Answer: 8 repeaters (9 pieces): 433 ps instead of 918 ps
Logical effort: sizing a path of gates for minimum delay
Which syllabus covers thisAU EC3552 Unit IIVTU BEC602 Module 3SPPU 404182 Unit IVMU not in syllabusKTU not in syllabus
How does logical effort express the delay of one gate?
Delay in units of τ is d = gh + p. h = Cout/Cin is the electrical effort, how much bigger the load is than the gate's own input. g, the logical effort, says how much worse than an inverter the gate is at driving current for its input capacitance. p is the parasitic delay, from the gate's own drain capacitance. For our illustrative process τ = 0.69 × 3.99 kΩ × 2.52 fF ≈ 6.9 ps. An inverter driving four copies of itself (FO4) takes 4 + 1 = 5τ ≈ 35 ps.
What are g and p for NAND and NOR gates with our mobility ratio of 2.5?
Size each gate to drive as well as the unit inverter: NMOS width 1, PMOS width 2.5, input capacitance 3.5. NAND2 has two NMOS in series, width 2 each, and parallel PMOS of 2.5, so g = 4.5/3.5 = 1.286. NOR2 has PMOS in series, 5 each, and NMOS of 1, so g = 6/3.5 = 1.714. NAND3 gives 1.571 and NOR3 2.429. Count the drain widths on the output and p = number of inputs. A process with ratio 2 gives the familiar 4/3 and 5/3. These values are ours.
How do you size a whole path of gates for the least delay?
Multiply the g values to get G, the branching factors to get B, and take H = Cout/Cin of the whole path: F = GBH. The delay is least when every stage bears the same effort f̂ = F^(1/N). Then work back from the load: each gate's input capacitance is g × (its load) / f̂. An AND3 (NAND3 then inverter) driving 50× its input has F = 78.6, f̂ = 8.86 and D = 21.7τ.
How many stages should a path have?
Add inverters and the stage effort falls, but each one costs its own parasitic delay. With p = 1 per inverter the best stage effort solves ρ(ln ρ − 1) = 1, giving ρ = 3.59. So N ≈ ln F / ln 3.6, close to the log₄ F rule of thumb. The optimum is flat: stage efforts from 2.4 to 6 cost under 9% extra. Each added inverter flips the polarity, so add an even number unless the logic can absorb the inversion.
A path has total effort F = 256 and only inverters can be added (p = 1 each). Which number of stages gives the least delay?
D = N·256^(1/N) + N gives 34τ for 2, 20τ for 4 and 24τ for 8 stages. Four stages put the stage effort at exactly 4, near the best value of 3.6.
Worked examples
Logical effort of 3-input gates from the transistor widths · Mobility ratio 2.5: unit inverter NMOS 1, PMOS 2.5 (input 3.5)
- NAND3: three NMOS in series, each width 3 (three in series behave like one of width 1); PMOS in parallel, each 2.5
- g(NAND3) = (3 + 2.5)/3.5 = 1.5714; output drains 3 × 2.5 + 3 = 10.5 = 3 × 3.5, so p = 3
- NOR3: three PMOS in series, each 7.5; NMOS in parallel, each 1
- g(NOR3) = (1 + 7.5)/3.5 = 2.4286; output drains 7.5 + 3 × 1 = 10.5, so p = 3
- With a ratio of 2 instead the same counting gives 1.6667 and 2.3333: NOR is hurt more by weak PMOS
Answer: g(NAND3) = 1.571, g(NOR3) = 2.429, p = 3 for both
Size an AND3 driving a large load · NAND3 then inverter, load = 50× the path's input capacitance
- G = 1.5714 × 1, H = 50, B = 1: F = 78.57
- Two stages: f̂ = √F = 8.864, D = 2 × 8.864 + (3 + 1) = 21.73τ ≈ 151 ps
- Inverter input cap = 1 × 50 / 8.864 = 5.64 (the NAND3 input is 1)
- One more inverter: D = 3·F^(1/3) + 5 = 17.85τ, the fastest, but the output becomes NAND3
- Two more inverters keep AND3: f̂ = 2.977, D = 17.91τ; caps from the input: 1, 1.89, 5.64, 16.79
Answer: As two gates: 21.7τ. With two buffer inverters: 17.9τ ≈ 124 ps
A path with branching · OR4 as NOR2 then NAND2; the NOR2 output drives three identical NAND2 inputs (b = 3); H = 20
- G = 1.7143 × 1.2857 = 2.2041, B = 3, F = GBH = 132.24
- f̂ = √F = 11.5, D = 2f̂ + 2 + 2 = 27τ
- NAND2 input = 1.2857 × 20 / 11.5 = 2.236; NOR2 input = 1.7143 × 3 × 2.236 / 11.5 = 1 ✓
- A numeric search over the NAND2 size finds the same minimum, 27τ at 2.236
- Two inverters at the end (polarity kept): D = 4·F^(1/4) + 6 = 19.56τ
Answer: 27τ as two gates; 19.6τ ≈ 136 ps with two added inverters
Pass-transistor logic and transmission gates
Which syllabus covers thisAU EC3552 Unit IIVTU BEC602 Modules 2, 4SPPU 404182 Unit IVMU ECC503 Module 3KTU not in syllabus
Why does an NMOS pass a strong 0 but a weak 1?
An NMOS conducts only while its gate is a threshold above its source. Passing a 0, the source is the low side and stays at 0 V, so the device stays fully on and pulls the output to 0. Passing a 1, the output is the source. As it rises, the overdrive shrinks and the body effect raises VT. Conduction stops when Vout = VDD − VTn(Vout): 1.106 V with our illustrative parameters, not 1.8 V. A PMOS is the mirror image: a strong 1 but a weak 0 of 0.694 V.
How does a transmission gate fix the threshold drop?
Put an NMOS (gate at VDD) and a PMOS (gate at 0) in parallel. Near 0 V the NMOS does the work, near VDD the PMOS, so the pair passes both levels to the rail. The cost is two transistors and a complementary control signal. Into 20 fF, a transmission gate passing a 1 reaches 90% of 1.8 V in 123 ps. A lone NMOS needs 626 ps just to reach 90% of its weak 1.11 V, because its overdrive dies away as the output rises.
How are a multiplexer and an XOR built from pass gates?
A 2:1 multiplexer is two transmission gates sharing the output, one turned on by S and the other by its complement. That is 4 transistors plus an inverter for the complement of S: 6 in all. A static CMOS version from an AOI22, an output inverter and a select inverter needs 12. An XOR passes A when B = 0 and the complement of A when B = 1: the same two-gate selector steered by B. Pass logic saves transistors, but every stage adds series resistance.
Do threshold drops add up along a chain of pass transistors?
Not when every gate sits at VDD. Each NMOS stops at VDD − VT(Vout), whatever its input, so four in series still end at 1.106 V, only more slowly. The drops do add when a pass output drives another transistor's gate: that second NMOS stops at 1.106 − VT, which is 0.506 V. A weak 1 driving an inverter also leaves its PMOS partly on, wasting static current. A level-restoring PMOS fed back from the inverter output pulls the node up to VDD and fixes this.
Four NMOS pass transistors in series, all with their gates at VDD = 1.8 V, pass a 1 into a capacitor. Where does the output settle?
Each transistor stops when its own source (the side nearer the output) reaches VDD − VTn of that node, so the first internal node settles at 1.11 V. The next one sees the same gate voltage and the same limit, so nothing more is lost. Only the time to get there grows.
Worked examples
The weak 1 of an NMOS pass transistor, including the body effect · Gate at VDD = 1.8 V, VT0 = 0.5 V, γ = 0.4 V^½, 2φF = 0.8 V, body at 0 V
- Final level V satisfies V = 1.8 − VT(V), with VT(V) = 0.5 + 0.4(√(0.8 + V) − √0.8)
- Guess V = 1.3 V (no body effect): VT = 0.7219 V, so V = 1.0781 V
- Repeat: V = 1.1096, 1.105, … converging to 1.1056 V (bisection agrees)
- Lost against the rail: 694 mV, of which 194 mV is body effect
- PMOS passing a 0 (body at VDD): by symmetry V = 0.6944 V
Answer: The NMOS passes a 1 as 1.106 V; the PMOS passes a 0 as 0.694 V
Series pass transistors versus a pass output driving a gate · Same process; case 1: two NMOS in series, gates at VDD; case 2: the output of one NMOS drives the gate of a second NMOS whose drain is at VDD
- Case 1: the first internal node settles at 1.106 V; the second NMOS also has gate 1.8 V and stops at the same 1.106 V
- Case 2: the second NMOS has only 1.106 V on its gate, so its output solves V = 1.106 − VT(V)
- Bisection: V = 0.506 V
- That 0.506 V is barely above VTn = 0.5 V: a later inverter would hardly see it as a 1
Answer: Case 1: 1.106 V (no extra loss); case 2: 0.506 V (the drops add)
Static current when a weak 1 drives an inverter · A pass-transistor output at 1.106 V drives our unit inverter (k'W/L = 400 µA/V² for both devices)
- PMOS: VSG = 1.8 − 1.106 = 0.694 V > |VTp| = 0.5 V, so it stays on with 194 mV overdrive
- With the output low the PMOS is saturated: I = ½ × 400 µA/V² × (0.1944)² = 7.56 µA
- The NMOS (overdrive 0.606 V, linear) must carry the same current: Vout = 32.1 mV
- Static power = 1.8 V × 7.56 µA = 13.6 µW, drawn for as long as the input sits there
Answer: The output is still a good 0 (32 mV), but 7.6 µA flows all the time: 13.6 µW
Pseudo-NMOS, dynamic and domino logic, and charge sharing
Which syllabus covers thisAU EC3552 Unit IIVTU BEC602 Modules 3-4SPPU not in syllabusMU ECC503 Module 3KTU not in syllabus
What is pseudo-NMOS logic and what does it cost?
Replace the whole PMOS pull-up network with one PMOS whose gate is tied to ground. An n-input gate then needs n + 1 transistors, not 2n. The catch is that the load is always on. With an input high, the NMOS network fights the PMOS, so the output settles at a ratioed VOL above 0 V, and current flows the whole time. With Wp/Wn = 0.5 a NOR2 with one input high sits at 0.137 V and draws 67.6 µA: 122 µW per gate, all from our illustrative parameters.
How does a dynamic gate work?
A clocked PMOS precharges the output to VDD while the clock is low; the foot NMOS is off, so no current flows. When the clock goes high (evaluate), the PMOS turns off and the foot turns on. If the NMOS network conducts, the output discharges; if not, it floats high on its capacitance. That needs n + 2 transistors, and the inputs drive only NMOS gates, so it is fast. But it can only fall during evaluate. Feed it into another dynamic gate and the second may discharge before the first has settled.
Why does domino logic add an inverter after each dynamic gate?
The inverter turns the precharged-high node into an output that is low during precharge and can only rise during evaluate. Every domino output starts at 0, so the next gate's inputs can only make a 0 → 1 change, and a cascade falls like a row of dominoes. The price: only non-inverting functions (AND, OR, AND-OR), and a weak keeper PMOS is usually added to hold the floating node high against leakage. Our inverter switches at VM = 0.9 V, which sets how far the node may droop.
What is charge sharing in a dynamic gate?
In a dynamic NAND2, let the top input A be 1 and B be 0 during evaluate. The output should stay high, but A connects it to the uncharged internal node X, and the charge spreads over both capacitances. The ideal formula is below. The droop also stops when A cuts off, at VX = 1.11 V. With X at 0 V, the output falls below VM = 0.9 V once Cx exceeds Cout, and the domino output falsely goes high.
A domino AND2 has Cx = 1.5 × Cout and its internal node starts at 0 V. In evaluate, A = 1 and B = 0. What does the domino output do?
Charge sharing drops the dynamic node to 1.8 / 2.5 = 0.72 V, below the inverter's switching threshold of 0.9 V, so the inverter output swings high: a false 1 that domino cannot undo until the next precharge.
Worked examples
VOL and static power of a pseudo-NMOS NOR2 · PMOS load W = 0.18 µm (Wp/Wn = 0.5), NMOS pull-downs Wn = 0.36 µm each, L = 0.18 µm
- PMOS (gate at 0) is saturated while VOL < 0.5 V: I = ½ × 80 µA/V² × 1 × 1.3² = 67.6 µA
- One NMOS on (linear): 400 µA/V² × (1.3 VOL − VOL²/2) = 67.6 µA, so VOL = 1.3(1 − √(1 − 0.4 × 0.5)) = 0.1372 V
- Both NMOS on: VOL = 1.3(1 − √(1 − 0.2 × 0.5)) = 0.0667 V
- Static power while low: 1.8 V × 67.6 µA = 121.7 µW
- A wider load rises faster but lifts VOL: at Wp/Wn = 1.55 VOL reaches VTn = 0.5 V and the next NMOS starts to conduct
Answer: VOL = 0.137 V (one input) or 0.067 V (both); 122 µW whenever the output is low
Charge sharing, with and without the transistor cutting off · Dynamic NAND2, Cout = 12 fF; X starts at 0 V; A = 1, B = 0 in evaluate; (a) Cx = 5 fF, (b) Cx = 15 fF
- (a) The simple formula gives 1.8 × 12/17 = 1.271 V, but X can only rise to 1.106 V, where A (gate at VDD) turns off
- Equalising would need X at 1.271 V > 1.106 V, so the transfer stops early: Vout = 1.8 − (5/12) × 1.106 = 1.339 V
- (b) Now 1.8 × 12/27 = 0.8 V is below 1.106 V, so the nodes do equalise: Vout = 0.8 V
- An RK4 simulation of the A transistor moving charge between the two capacitors settles at the same values
- (b) is below VM = 0.9 V: a domino inverter would switch
Answer: (a) 1.34 V: safe; (b) 0.8 V: the domino output falsely goes high
How large may the internal capacitance be? · Cout = 12 fF, next stage switches at VM = 0.9 V
- With X at 0 V, the droop crosses 0.9 V where 1.8/(1 + r) = 0.9 (0.9 V is below the 1.11 V cut-off, so the equalising branch applies)
- r = Cx/Cout = 1: Cx may be at most 12 fF
- If X is precharged too (an extra small PMOS, or left at 1.11 V from a previous cycle), A is already cut off and the output stays at 1.8 V
- Otherwise keep X small (little shared diffusion and wiring) or add a keeper PMOS that tops the output back up
Answer: Cx ≤ 12 fF with X starting at 0 V; no limit if X is precharged
Latches, flip-flops and register timing: setup, hold and clock-to-Q
Which syllabus covers thisAU EC3552 Unit IIIVTU BEC602 Module 5SPPU 404182 Unit IIMU ECC503 Module 3KTU PEECT415 Module 3
What is the difference between a latch and a flip-flop?
A latch is level-sensitive. A positive latch is transparent while the clock is high, so its output follows the input, and it holds the last value while the clock is low. A flip-flop (an edge-triggered register) is built from two latches in series, a negative master and a positive slave. The master is open while the clock is low and the slave while it is high, so the output changes only just after the rising edge. Never both open at once: that is what stops data racing straight through.
What is the difference between a static and a dynamic register?
A static latch holds its value with feedback: two cross-coupled inverters, with transmission gates or clocked inverters choosing between the input and the loop. It keeps its state as long as power stays on, even with the clock stopped. A dynamic latch drops the feedback and stores the bit as charge on a node capacitance, which saves transistors and is faster. Leakage drains it, though. A 4 fF node losing 10 pA drops 0.3 V in 120 µs, so the clock must keep refreshing it: above about 4.17 kHz if it must hold for half a period (illustrative values).
What is a C2MOS register and why does it tolerate clock overlap?
C2MOS (clocked CMOS) builds each latch from a clocked inverter: a PMOS pair in series above an NMOS pair, with one transistor of each pair driven by the clock or its complement. A master-slave C2MOS register uses two such stages with opposite clock phases. If both clocks are briefly low, only the clocked PMOS devices conduct, so each stage can pull its output up but never down. The master can only make its output rise, and the inverting slave would then have to fall, which needs its clocked-off pull-down. So new data cannot race through. Both clocks high is the mirror case. This holds while clock rise and fall times stay short.
What are clock-to-Q, setup and hold time, and how do you check a register-to-register path?
Clock-to-Q (tcq) is how long a flip-flop's output takes to change after the clock edge. Setup time (tsu) is how long before the edge the input must be stable. Hold time (th) is how long after the edge it must stay stable. On a path the slowest logic (tpd) sets the setup check, and the fastest logic (tcd) sets the hold check. Take a period of 800 ps, tcq = 70 ps, tpd = 600 ps and tsu = 50 ps: setup slack = 800 − 720 = 80 ps. With tcd = 40 ps and th = 20 ps, hold slack = 70 + 40 − 20 = 90 ps.
How does clock skew change the setup and hold checks, and how do you fix a hold violation?
This book's sign convention: skew δ = capture clock arrival − launch clock arrival, so positive skew means the capture clock is late. Late capture gives the slow path more time, which helps setup, but lets new data reach a register still holding the old value, which hurts hold. In the 800 ps example, δ = +100 ps raises setup slack to 180 ps and drops hold slack to −10 ps. Slowing the clock never fixes hold, because the hold check does not depend on the period. The fix is to add delay (buffers) on the short path.
A register path passes both checks. The clock now reaches the capturing flip-flop 100 ps later than the launching one (positive skew). Which check is at risk?
A late capture clock lengthens the time the slow path has (setup slack rises by 100 ps) but gives the fast path 100 ps longer to overrun the old data before the capture register finishes holding it. In the 800 ps example, hold slack goes from +90 to −10 ps, a violation that no change of clock period can fix.
Worked examples
Maximum clock frequency of a register path, with and without skew · tcq = 80 ps, slowest logic tpd = 620 ps, setup time tsu = 45 ps; first zero skew, then the capture clock 70 ps late (δ = +70 ps)
- Setup needs T ≥ tcq + tpd + tsu − δ. With δ = 0: Tmin = 80 + 620 + 45 = 745 ps
- fmax = 1/745 ps = 1.342 GHz
- With δ = +70 ps the capture edge arrives later, so Tmin = 745 − 70 = 675 ps and fmax = 1.481 GHz
- A grid check of the data at the capture input (launch every T, data unknown while the logic settles) passes at 1 ps above Tmin and fails 1 ps below it
Answer: 1.342 GHz with zero skew, 1.481 GHz with 70 ps of positive skew, if hold still passes (next example)
The same skew breaks hold, and a buffer fixes it · The path above also has a fastest logic path tcd = 30 ps, and the capture flip-flop needs th = 50 ps; skew δ = +70 ps; a delay buffer adds 15 ps
- Hold needs tcq + tcd ≥ th + δ. With δ = 0: slack = 80 + 30 − 50 = +60 ps
- With δ = +70 ps: slack = 80 + 30 − 50 − 70 = −10 ps: the fast data can overrun the old value at the capture register
- Slowing the clock does not help: T does not appear in the hold check
- Put one 15 ps buffer on the short path only: tcd = 45 ps, slack = 80 + 45 − 50 − 70 = +5 ps
- The slow path is unchanged, so Tmin stays 675 ps
Answer: Hold slack −10 ps → +5 ps with one 15 ps buffer on the short path; fmax stays 1.481 GHz
How slow can a dynamic register be clocked? · Dynamic latch storage node 4 fF, leakage 10 pA, the next stage still reads a 1 correctly after a droop of up to 0.3 V; the node must hold for half a clock period (illustrative)
- Leakage drains the node at a constant rate: dV/dt = I/C = 10 pA / 4 fF = 2.5 V/ms
- Time to droop 0.3 V: t = C·ΔV/I = 4 fF × 0.3 V / 10 pA = 120 µs (a 1 µs step-by-step integration agrees)
- Half a period must be shorter than this: T/2 ≤ 120 µs, T ≤ 240 µs
- fmin = 1/240 µs = 4.17 kHz
Answer: Clock at 4.17 kHz or faster; a static register has no such minimum
Pipelining: throughput, latency and register overhead
Which syllabus covers thisAU EC3552 Unit IIIVTU not in syllabusSPPU 404182 Unit IMU not in syllabusKTU not in syllabus
What does pipelining buy you, and what does it cost?
Pipelining cuts a long block of logic into stages with a register between each, so a new input can start every clock instead of waiting for the previous one to finish. Throughput (results per second) goes up. Latency (time for one result) goes slightly up too, because every stage pays the register overhead. Take 8 ns of logic and a 0.4 ns overhead (tcq + tsetup). One stage: 8.4 ns per result, 119 MHz. Four stages: a 2.4 ns clock (417 MHz), 3.5× the throughput, but each result now takes 9.6 ns.
Why do more stages give diminishing returns?
Each stage's clock period is its logic delay plus the fixed register overhead. Splitting finer shrinks only the logic part, so the period can never drop below the overhead, and throughput can never beat 1/overhead. With 8 ns of logic and 0.4 ns overhead that ceiling is 2.5 results per ns. Going from 4 to 8 stages takes the period from 2.4 to 1.4 ns: only 1.71× the throughput for twice the registers, plus more clock load and power.
Why must pipeline stages be balanced, and why can't real logic be cut anywhere?
The clock has to suit the slowest stage, so any stage longer than the rest wastes time in all the others. With 8 ns over four stages, making one stage 1.3× the others stretches the period from 2.4 to 2.819 ns, 17.4% slower. Real logic can only be cut between gates. In the interactive's chain of 12 unequal gates, the best four-way cut has a longest stage of 2.2 ns, not 2, so the clock is 2.6 ns. Designers move logic across registers (retiming) to even this out.
How do latch-based pipelines differ from flip-flop pipelines?
Replace each flip-flop with two latches on opposite clock phases and put logic between every latch. A latch is transparent for half a cycle, so a slow block can finish late and borrow time from the next block, as long as each pair of blocks fits in a full cycle. With a 2 ns clock, a 1.2 ns block can borrow 0.2 ns if the block after it needs only 0.8 ns. Time borrowing smooths out imbalance and hides some clock skew. The cost is a harder timing analysis and a hold check on every latch.
A pipeline has 8 ns of logic in 4 balanced stages with 0.4 ns of register overhead per stage. You double it to 8 stages. What happens to throughput?
The period goes from 2 + 0.4 = 2.4 ns to 1 + 0.4 = 1.4 ns. Only the logic part halves; the 0.4 ns overhead is paid in every stage. Throughput rises 1.71×, and latency grows from 9.6 ns to 11.2 ns.
Worked examples
Clock, latency and throughput for 1, 3 and 6 stages · 12 ns of combinational logic, split evenly; register overhead tcq + tsetup = 0.5 ns per stage
- 1 stage: T = 12 + 0.5 = 12.5 ns, f = 80 MHz, latency 12.5 ns
- 3 stages: T = 12/3 + 0.5 = 4.5 ns, f = 222.2 MHz, latency 3 × 4.5 = 13.5 ns
- 6 stages: T = 12/6 + 0.5 = 2.5 ns, f = 400 MHz, latency 6 × 2.5 = 15 ns
- Speed-up over 1 stage = 12.5/T: 2.778× for 3 stages, 5× for 6 stages
Answer: 80, 222.2 and 400 MHz; latency rises from 12.5 to 15 ns
The cost of unbalanced stages · The same 12 ns of logic can only be cut at existing boundaries, giving natural stages of 3, 4 and 5 ns; overhead 0.5 ns
- The clock must suit the 5 ns stage: T = 5 + 0.5 = 5.5 ns, f = 181.8 MHz
- A perfectly balanced three-way split would give T = 4.5 ns, f = 222.2 MHz
- Throughput lost: 1 − 4.5/5.5 = 18.2%
- Moving 1 ns of logic from the third stage into the first (retiming) recovers the balanced figure
Answer: 181.8 MHz instead of 222.2 MHz, 18.2% less throughput
How many stages for a 4× speed-up? · 12 ns of logic split evenly, 0.5 ns register overhead; target: 4× the throughput of the unpipelined design (12.5 ns)
- Speed-up = 12.5/(12/k + 0.5) ≥ 4 means 12/k + 0.5 ≤ 3.125 ns
- 12/k ≤ 2.625, k ≥ 4.571, so k = 5
- Check by trying each k: k = 4 gives 3.571×, k = 5 gives 4.31×
- The ceiling as k grows without limit is 12.5/0.5 = 25×, never reached
Answer: 5 stages (speed-up 4.31×)
Clock generation and distribution: skew, jitter and the H-tree
Which syllabus covers thisAU EC3552 Unit IIIVTU not in syllabusSPPU 404182 Unit IIMU ECC503 Module 5KTU PEECT415 Module 3
Where does an on-chip clock come from, and what are skew and jitter?
A crystal outside the chip gives a stable but slow reference, and an on-chip phase-locked loop (PLL) multiplies it. A 25 MHz reference times 40 gives 1 GHz (illustrative). The clock then travels a network of wires and buffers to every flip-flop. Skew is a difference in arrival time between two flip-flops on the same edge. Jitter is the variation of one flip-flop's clock edges from cycle to cycle. Skew is mostly fixed by the layout and can be measured and balanced. Jitter is random and has to be budgeted.
How does an H-tree give zero skew, and what does it cost?
An H-tree drives the centre of an H whose four tips sit at the centres of the four quadrants. Each tip drives a half-size H, and so on, so every path from the root to a leaf has exactly the same length and the same mix of wires. With matched wires every leaf gets the clock at the same time. A 2 mm block with 3 levels has 64 leaves, each 1.75 mm of wire from the root, and 21 mm of wire in total. Equal delays cost wire: a spine reaching the same 64 leaves uses only 15.75 mm.
Why do real clock trees still have skew?
A tree that is not symmetric gives different delays to different leaves. In this chapter's illustrative model (25 Ω driver, 10 fF per leaf, no buffers, 2 mm block, 64 leaves), a spine is faster at worst (177.4 ps against 275.6 ps for the H-tree) but its leaves spread over 61.8 ps. Even an H-tree drifts when wires are not identical: 20% wider wires in one quadrant give 5.7 ps of skew, 20% narrower give 8.6 ps. Large chips add buffers at the branch points, tune them, and often tie the leaves into a grid that averages out the differences.
How do skew and jitter eat into the timing budget?
For setup, the worst case is a launch clock that is late and a capture clock that is early, plus jitter that shortens the cycle. Both come straight out of the time left for logic. At 1 GHz with tcq = 60 ps, tsu = 40 ps, 30 ps of skew against you and 20 ps of jitter, the logic gets 850 ps of the 1000 ps cycle. Hold compares the same clock edge at both registers, so cycle-to-cycle jitter mostly cancels, but skew in the hold direction adds straight to the hold requirement.
In an H-tree with perfectly matched wires, which flip-flops receive the clock first?
In an H-tree every leaf sits at the end of an identical path from the root: the same number of branches, the same wire lengths and the same loads. With matched wires the Elmore delays are equal and the nominal skew is zero, even though the leaves are at different distances from the centre as the crow flies.
Worked examples
Wire length of a four-level H-tree · Square block 3 mm on a side, H-tree with 4 levels; wire capacitance 0.2 fF/µm, 10 fF per leaf (illustrative)
- Leaves: 4⁴ = 256
- Level l has 4^(l−1) H's, each made of six arms a quarter of its size: wire at level l = 4^(l−1) × 6 × S/2^(l+1)
- Total = 1.5 × S × (2⁴ − 1) = 1.5 × 3 × 15 = 67.5 mm (a recursive build of the tree agrees)
- Root to any leaf: S(1 − 2⁻⁴) = 3 × 15/16 = 2.8125 mm, the same for all 256 leaves
- Wire capacitance 67.5 mm × 0.2 fF/µm = 13.5 pF, plus 2.56 pF of leaf loads
Answer: 67.5 mm of wire (13.5 pF), 2.8125 mm to every leaf
Elmore delay of one H, and the skew from one wider wire · 2 mm block, one H (4 leaves), driver 25 Ω, leaf load 10 fF, wire 0.08 Ω/µm and 0.2 fF/µm; each of the six arms is 500 µm
- Each arm: R = 0.08 × 500 = 40 Ω, C = 0.2 × 500 = 100 fF
- Total capacitance = 6 × 100 + 4 × 10 = 640 fF; driver term 25 Ω × 640 fF = 16 ps
- Centre arm: 40 Ω × (50 + 2 × 110) fF = 10.8 ps; leg to a leaf: 40 Ω × (50 + 10) fF = 2.4 ps
- Every leaf: 16 + 10.8 + 2.4 = 29.2 ps, skew 0
- Make the top-right leg 20% wider (R ÷ 1.2, C × 1.1): the leaves now see 29.45 to 29.85 ps, a skew of 0.4 ps
Answer: 29.2 ps to every leaf; one mismatched leg gives 0.4 ps of skew
Logic budget with skew and jitter · 1.25 GHz clock (800 ps), tcq = 55 ps, tsu = 35 ps, th = 20 ps; up to 40 ps of skew and 25 ps of cycle-to-cycle jitter
- Setup, worst case: the skew and the jitter both shorten the cycle the logic sees
- Logic ≤ 800 − 55 − 35 − 40 − 25 = 645 ps
- Hold, worst case: the capture clock 40 ps late; jitter cancels because both registers use the same edge
- tcq + tcd ≥ th + skew: 55 + tcd ≥ 20 + 40, so every short path needs tcd ≥ 5 ps
Answer: Slowest logic ≤ 645 ps; fastest logic ≥ 5 ps
Asynchronous inputs, metastability and synchronizers
Which syllabus covers thisAU EC3552 Unit IIIVTU not in syllabusSPPU 404182 Unit IIMU not in syllabusKTU PEECT415 Module 3
What is metastability?
A flip-flop stores a bit in two cross-coupled inverters, which have two stable states and one balance point in between, at the switching threshold VM. If the input changes just as the clock closes the latch, the node can be left almost at VM, like a ball balanced on a hilltop. The loop amplifies any offset exponentially, V − VM = ΔV0·e^(t/τ), so it always falls one way eventually, but how long it takes depends on how close it started. Every 10× closer to VM costs τ·ln 10: 46 ps with τ = 20 ps.
How is a synchronizer's mean time between failures (MTBF) calculated?
An input edge landing inside a small window T0 around the clock edge leaves the flip-flop metastable, which happens T0·fclk·fdata times a second. A failure is one still unresolved after the resolution time S, which has probability e^(−S/τ). So MTBF = e^(S/τ)/(T0·fclk·fdata) (Ginosar, IEEE Design & Test, 2011). The exponent dominates. With τ = 20 ps, T0 = 30 ps, 100 MHz data and two flip-flops, a 1 GHz clock leaves S = 920 ps and an MTBF of about 1.0 × 10⁶ years. At 2 GHz, S = 420 ps and the MTBF falls to 220 seconds.
How do you bring a signal safely into a clock domain?
Pass a single-bit signal through two or more flip-flops in series, with no logic between them. The first may go metastable, and the rest of the cycle is its time to resolve before the second samples it. Each extra flip-flop adds roughly one more cycle of resolution and multiplies the MTBF by e^((Tclk − toh)/τ). Never synchronize the bits of a bus one by one, because they can resolve differently. Send a Gray-coded count (one bit changes at a time), a handshake, or use an asynchronous FIFO. Ginosar's example (τ = 10 ps, T0 = 20 ps, 1 GHz, data every 10 cycles, one cycle to resolve) works out to about 4 × 10²⁹ years.
What is self-timed (asynchronous) design?
A self-timed circuit has no global clock. Each block tells the next when its data is ready with a request, and the receiver answers with an acknowledge: a handshake. Bundled-data designs send one wire per bit plus a request delayed to match the logic. Dual-rail designs use two wires per bit, so the data itself shows when it has arrived (delay-insensitive). A 32-bit bus then needs 64 wires. The gains are power only where work happens, average-case rather than worst-case delay, and no clock tree. The costs are handshake area and harder tools. Arbiters, which decide which request came first, must wait out metastability.
A two-flip-flop synchronizer at 2 GHz (τ = 20 ps, 80 ps overhead per hop) fails about every 220 seconds. You add a third flip-flop. Roughly how much longer is the MTBF?
The third flip-flop adds one more hop of resolution time, 500 − 80 = 420 ps, and the MTBF grows as e^(S/τ): e^(420/20) = e^21 ≈ 1.3 × 10⁹. The first flip-flop still goes metastable just as often, but the chance it is still undecided after two hops is vanishingly small: the MTBF goes from 220 s to about 9,186 years.
Worked examples
Two or three flip-flops at 2 GHz? · Clock 2 GHz (500 ps), data changes at 100 MHz, τ = 20 ps, T0 = 30 ps, tcq + tsetup = 80 ps per hop (illustrative)
- Two flip-flops: S = 500 − 80 = 420 ps, S/τ = 21
- T0·fclk·fdata = 30 ps × 2 GHz × 100 MHz = 6 × 10⁶ per second
- MTBF = e^21/6 × 10⁶ = 219.8 s: a failure every few minutes
- Three flip-flops: S = 2 × 420 = 840 ps, MTBF = e^42/6 × 10⁶ = 2.90 × 10¹¹ s = 9,186 years
Answer: Two flip-flops: 220 s, unusable. Three: about 9,186 years
How small must τ be? · 1 GHz clock, data at 50 MHz, T0 = 25 ps, two flip-flops with 80 ps overhead; target MTBF 100 years
- S = 1000 − 80 = 920 ps
- 100 years = 3.156 × 10⁹ s; need e^(S/τ) ≥ 3.156 × 10⁹ × 25 ps × 1 GHz × 50 MHz = 3.945 × 10¹⁵
- S/τ ≥ ln(3.945 × 10¹⁵) = 35.91
- τ ≤ 920/35.91 = 25.62 ps (bisection on the MTBF itself agrees)
Answer: τ ≤ 25.6 ps; a slower process needs a third flip-flop
How often does the first flip-flop go metastable? · 1 GHz clock, data at 50 MHz arriving at random times, T0 = 25 ps, τ = 20 ps; resolution time 920 ps in a two-flip-flop synchronizer
- A data edge falls inside the 25 ps window with probability T0·fclk = 25 ps/1 ns = 0.025 (a random-time simulation of 400,000 edges agrees within 5%)
- So the first flip-flop goes metastable 1.25 million times a second
- Probability one of those is still unresolved 920 ps later: e^(−920/20) = e^−46 = 1.1 × 10⁻²⁰
- Failure probability per data edge = 0.025 × 1.1 × 10⁻²⁰ = 2.6 × 10⁻²²; failure rate 1.3 × 10⁻¹⁴ per second
Answer: Metastable 1.25 million times a second, yet a failure probability of only 2.6 × 10⁻²² per edge
Fast adders: carry-lookahead, carry-select, carry-skip and carry-save
Which syllabus covers thisAU EC3552 Unit IVVTU not in syllabusSPPU not in syllabusMU ECC503 Modules 3, 5KTU not in syllabus
Why is a ripple-carry adder slow, and what are generate and propagate?
Each bit of a ripple adder waits for the carry from the bit below, so the delay grows with the width. In this chapter's gate model (AND and OR = 1 gate delay, XOR and a 2:1 mux = 2) an n-bit ripple adder needs 2n + 2 gate delays: 34 at 16 bits, 130 at 64. Every fast adder starts from two signals per bit. Generate g = ab: this bit makes a carry by itself. Propagate p = a XOR b: this bit passes an incoming carry on. Both are ready after 2 gate delays, for all bits at once.
How does a carry-lookahead adder compute carries without waiting?
Substitute the carry relation into itself: c2 = g1 + p1g0 + p1p0c0, and so on. Every carry becomes a two-level AND-OR of g, p and the carry-in. Wide gates get slow, so the expansion is cut into 4-bit blocks. Each block reports a group generate G and a group propagate P, and a lookahead unit treats the blocks like bits. With one level (block carries rippling block to block) a 16-bit adder takes 13 gate delays. With a second level over groups of four blocks it takes 10, and at 64 bits 17 instead of 130.
How do carry-select and carry-skip adders save time?
Carry-select builds each block twice, once assuming carry-in 0 and once 1, and a mux picks the right result when the real carry arrives: about double the area, 16 gate delays at 16 bits with 4-bit blocks. Carry-skip keeps one ripple chain per block but adds a mux that lets a carry jump a block whose bits all propagate. With block size k the worst case is 4k + 2n/k gate delays here, smallest at k = √(n/2): 32 at 32 bits with 4-bit blocks. A static timer that ignores logic values still sees the ripple through every block, a false path.
What is a carry-save adder and where is it used?
A carry-save adder takes three numbers and returns two, a sum word S and a carry word K, using one full adder per bit and no carry chain at all. S is ready after 4 gate delays whatever the width. One ordinary adder then adds S and K shifted left by one. For three 8-bit numbers 0x5B + 0x37 + 0x6E, S = 0x02 and K = 0x7F, and S + 2K = 256. With only three operands the saving is small. It pays off when many numbers are added, as inside a multiplier.
A 64-bit ripple adder takes 130 gate delays in this chapter's model. You build it as a carry-skip adder with the best block size. About how many gate delays now?
The worst case is 4k + 2n/k. At n = 64, 4-bit and 8-bit blocks both give 48 gate delays: the carry ripples through the first block, skips the middle blocks two gate delays each, and ripples through the last block. The delay grows like √n, not n, so doubling the width multiplies it by about 1.4.
Worked examples
Generate, propagate and lookahead carries for a 4-bit add · A = 1011 (11), B = 0110 (6), carry-in 0
- From bit 0 up: g = ab = 0, 1, 0, 0; p = a XOR b = 1, 0, 1, 1
- c1 = g0 + p0c0 = 0; c2 = g1 + p1g0 = 1; c3 = g2 + p2g1 + p2p1g0 = 1
- c4 = g3 + p3g2 + p3p2g1 + p3p2p1g0 = 1: this is the block's group generate G
- Group propagate P = p3p2p1p0 = 0 (bit 1 does not propagate, it generates)
- Sum bits si = pi XOR ci give 0001 with carry-out 1
Answer: A + B = 10001 = 17; all four carries come from two gate levels after g and p, none waits for the bit below
Delay of a 16-bit adder in five architectures · Gate model: AND/OR (up to 4 inputs) = 1, XOR and 2:1 mux = 2; inputs at t = 0
- Ripple: each bit adds an AND and an OR to the carry path: 2 × 16 + 2 = 34
- CLA, one level: g, p (2), G of each block (4), block carries 2 apart, last block's carries and sums: 13
- CLA, two levels: the four block carries come straight from the second-level unit: 10
- Carry-select, 4-bit blocks: the first block ripples (10), then one mux (2) per block: 16
- Carry-skip, 4-bit blocks: ripple out of block 0, skip blocks 1 and 2, ripple through block 3: 24
Answer: ripple 34, CLA 13, two-level CLA 10, carry-select 16, carry-skip 24 gate delays
Choosing the block size of a 32-bit carry-skip adder · n = 32, uniform blocks of k bits, delay 4k + 2n/k gate delays (k < n)
- k = 1: 4 + 64 = 68; k = 2: 8 + 32 = 40
- k = 4: 16 + 16 = 32; k = 8: 32 + 8 = 40; k = 16: 64 + 4 = 68
- Setting d(4k + 64/k)/dk = 0 gives k = √(n/2) = 4
- A floating-mode simulation of the worst operands (a carry made in bit 0 that must run to the top) reproduces each value
- A static timer that ignores values reports 82 for k = 4: the ripple through every block, which no input can trigger
Answer: 4-bit blocks: 32 gate delays against 66 for ripple
Multipliers: array, carry-save and Booth recoding
Which syllabus covers thisAU EC3552 Unit IVVTU not in syllabusSPPU not in syllabusMU ECC503 Module 5KTU not in syllabus
How does an array multiplier work and how slow is it?
Multiplying two n-bit numbers makes n partial products: the multiplicand ANDed with each bit of the multiplier, each shifted one place further left. That is n² AND gates, all switching at once. An array multiplier adds the rows one at a time in carry-save form, so no carry chain forms inside the array, and a final ripple adder merges the last sum and carry words. For 8 × 8 bits this chapter's array uses 48 full adders and 8 half adders. The rows take 7 adder delays and the whole multiply 14: about 2n.
How do you multiply signed (two's complement) numbers in an array?
In two's complement the top bit carries a negative weight, so partial products that pair a sign bit with a normal bit are negative. The Baugh-Wooley trick keeps every row positive. Those mixed products are complemented, and two constant 1s are added, at column n and at column 2n − 1. The sign-bit times sign-bit product stays as it is. The array is then exactly the unsigned one with a few inverters and two extra dots. The product is read as a 2n-bit two's complement number, and every 4-bit pair checks out in this book's tests.
What does radix-4 Booth recoding do?
Booth's idea (Andrew Booth, 1950) is to recode the multiplier so runs of 1s cost almost nothing. The radix-4 version looks at overlapping groups of three multiplier bits and turns each into one digit from −2 to +2. Each digit needs only the multiplicand, its double (a shift), its negative or zero, so there is no real multiply. An n-bit signed multiplier gives n/2 digits instead of n bits, halving the partial products: 4 for 8 bits, 8 for 16. An unsigned multiplier needs one extra digit, ⌈(n + 1)/2⌉.
What do Wallace and Dadda trees change?
Instead of adding the rows one after another, a tree squeezes all of them at once. Every full adder turns three bits of one weight into two. Wallace's tree (1964) does as much as it can at each stage. Dadda's (1965) does as little as it can while keeping the stage count minimal, using heights 2, 3, 4, 6, 9, 13, 19, 28, 42, 63. The number of stages grows like a logarithm: 4 for 8 bits, 6 for 16, 8 for 32, 10 for 64. A tree only pays off with a fast final adder. With a ripple one, the 8 × 8 tree still takes 14 adder delays, like the array.
A 16 × 16 signed multiplier uses radix-4 Booth recoding. How many partial products are left to add?
Each radix-4 digit covers two multiplier bits, so a 16-bit signed multiplier gives 16/2 = 8 digits, each choosing 0, ±X or ±2X. The tree then reduces about 8 rows instead of 16 (sign handling adds a few bits): by the Dadda heights, 4 stages instead of 6.
Worked examples
Radix-4 Booth recoding of a negative multiplier · X = 53, Y = −75, both 8-bit two's complement (Y = 10110101)
- Append y−1 = 0 and read the bits in overlapping triples (y2i+1, y2i, y2i−1), two bits apart
- i = 0: (010) → d = 1; i = 1: (010) → d = 1; i = 2: (110) → d = −1; i = 3: (101) → d = −1
- Check the recoding: 1 + 1 × 4 − 1 × 16 − 1 × 64 = −75
- Partial products d·X·4^i: 53, 212, −848, −3392
- Sum: 53 + 212 − 848 − 3392 = −3975
Answer: Four partial products instead of eight, and 53 × (−75) = −3975
Hardware and delay of an 8 × 8 unsigned multiplier · Array (rows added one at a time in carry-save form) against a Dadda tree, both with a ripple final adder; 1 adder = 1 FA delay
- Partial products: 8 × 8 = 64 AND gates, 15 columns of dots, the tallest 8 high
- Array: rows 1 to 7 each add one partial product: 48 FAs and 8 HAs in all, the last row finishes after 7 adder delays
- Dadda: heights 8 → 6 → 4 → 3 → 2, so 4 stages with the same 48 FAs and 8 HAs
- Both end with a ripple adder along the 2 × 8 columns, so both finish after 14 adder delays
- Each full adder removes exactly one dot, so the full-adder count is fixed by the dot count; the half adders and the wiring differ
Answer: 48 FAs + 8 HAs either way; reduction takes 7 delays in the array against 4 in the tree, but 14 in total until the final adder is fast
A signed 4 × 4 product with Baugh-Wooley · x = −5 (1011), y = 6 (0110), 4-bit two's complement
- Partial-product rows (bit 3 to bit 0), with x_i·y_3 (i < 3) and x_3·y_j (j < 3) complemented: 1000, 0011, 0011, 0111
- Shift row j left by j and add: 82
- Add the two constant 1s at column 4 (16) and column 7 (128): 82 + 16 + 128 = 226
- Keep 8 bits: 226 mod 256 = 226 = 11100010
- Read as 8-bit two's complement: 226 − 256 = −30
Answer: −5 × 6 = −30, from the plain unsigned array plus inverters and two 1s
Memory architecture and the 6T SRAM cell: read and write stability
Which syllabus covers thisAU EC3552 Unit IVVTU not in syllabusSPPU not in syllabusMU ECC503 Module 4KTU not in syllabus
How is a memory array organised?
A memory is a grid of cells. Each row shares a wordline and each column shares a pair of bitlines. The row decoder turns part of the address into one active wordline. The column decoder drives a mux that picks which bitlines reach the outputs. Making the array nearly square keeps both kinds of line short. A 64 Kib block as 256 rows × 256 columns with a 4:1 column mux delivers 64-bit words: 8 address bits pick the row and 2 pick the column group. A sense amplifier on each selected column detects a small bitline swing (0.1 V in this chapter's illustrative numbers), so a read need not wait for a full swing.
How does the 6T SRAM cell hold, read and write a bit?
Two cross-coupled inverters hold the bit as Q and its complement QB, each node keeping the other in place for as long as power stays on. Two NMOS access transistors connect Q and QB to the bitlines BL and BLB when the wordline is high. To read, both bitlines are precharged high and the wordline opens. The side holding 0 slowly pulls its bitline down and the sense amplifier sees the difference. To write, the drivers force one bitline to 0 and the cell is overpowered through the access transistor.
Why can reading upset an SRAM cell, and what is the cell ratio?
During a read the '0' node sits between a pull-down to ground and an access transistor fed from a bitline at VDD, so it rises by ΔV. If ΔV gets near the other inverter's threshold the cell can flip. The cell ratio CR = (W/L) of the pull-down over (W/L) of the access sets the bump. In this book's illustrative long-channel process, ΔV = 0.381 V at CR = 1 and 0.239 V at CR = 2. It stays below VTn = 0.5 V for CR above 0.61. The static noise margin shows the cost: 297 mV during a read at CR = 1.5, against 678 mV when just holding.
What limits writability, and what is the pull-up ratio?
To write a 0 into a node holding 1, the access transistor (bitline at 0) must pull the node down against the PMOS pull-up that holds it high. The pull-up ratio PR = (W/L) of the pull-up over (W/L) of the access sets the fight. The node must fall below VTn so the other pull-down lets go. Here it falls to 0.293 V at PR = 1, writable, and reaches 0.5 V at PR = 1.55. Reading wants the pull-down strong against the access transistor and writing wants the pull-up weak against it, so the pull-down is the widest of the three.
To read faster you make both access transistors of an SRAM cell wider, leaving the inverters alone. What happens to the read bump on the '0' node?
A wider access transistor lowers the cell ratio CR, so the access wins more of its fight with the pull-down and the '0' node rises further: from 0.239 V at CR = 2 to 0.381 V at CR = 1 in this book's process. The read is faster, but stability is lower. Writing does get easier, since PR also falls.
Worked examples
Read bump at a cell ratio of 2 · Illustrative process VDD = 1.8 V, VTn = 0.5 V, k'n = 200 µA/V²; access W/L = 1, pull-down W/L = 2; bitline at VDD; Q stores 0, QB = VDD
- Access: gate at VDD, drain at VDD, source at ΔV, so VDS = VGS and it is saturated: I = ½ k'n (1.3 − ΔV)²
- Pull-down: gate at VDD, VDS = ΔV small, linear: I = 2 k'n (1.3 ΔV − ΔV²/2)
- Equate: (1.3 − ΔV)² = 4(1.3 ΔV − ΔV²/2), a quadratic with root ΔV = 1.3(1 − √(2/3))
- ΔV = 0.2386 V, below VTn = 0.5 V; bisection on the two currents gives the same value
- Read current at that point: 112.7 µA through both transistors
Answer: ΔV = 0.239 V; the cell reads safely (the limit ΔV = VTn is reached at CR = 0.61)
Write level and the largest pull-up · Same process, k'p = 80 µA/V², |VTp| = 0.5 V; Q holds 1, BL driven to 0, QB still at 0 (the start of the write)
- Access: gate VDD, source at BL = 0, drain at Q = V: linear, I = k'n (1.3 V − V²/2)
- Pull-up: gate at QB = 0, VSG = 1.8 V, VSD = 1.8 − V; saturated while V ≤ 0.5 V
- PR = 1: the access is linear and the pull-up saturated at the balance (½ × 80 × 1.3² = 67.6 µA), V = 0.293 V < 0.5 V, so the write succeeds
- PR = 1.5: V = 0.478 V, just writable
- Setting V = VTn with the pull-up saturated: k'n (1.3 × 0.5 − 0.125) = ½ k'p PR × 1.3², so PR = 1.553
Answer: V = 0.293 V at PR = 1; writes fail (by this conservative test) above PR ≈ 1.55
Organising a 64 Kib block and timing its bitline · 65,536 cells as 256 rows × 256 columns, 4:1 column mux; illustrative bitline 250 fF; read current from the previous example
- Word width: 256 columns ÷ 4 = 64 bits; words: 65,536 ÷ 64 = 1024
- Address: log2 1024 = 10 bits, 8 for the row decoder (256 wordlines) and 2 for the column mux
- A sense amplifier needs a 0.1 V swing: t = C ΔV / I = 250 fF × 0.1 V ÷ 112.7 µA = 222 ps
- A full 1.8 V swing at the same current would take 18 times longer, which is why SRAMs sense small swings
- The square-law model overstates the current of a real minimum cell, so treat the time as illustrative
Answer: 1024 words of 64 bits, 10 address bits; about 222 ps to develop 0.1 V
DRAM, ROM and flash memory cells
Which syllabus covers thisAU EC3552 Unit IVVTU not in syllabusSPPU not in syllabusMU ECC503 Module 4KTU not in syllabus
How does a one-transistor DRAM cell store and read a bit?
The 1T1C cell (Robert Dennard, IBM, 1966) stores a bit as charge on a small capacitor Cs behind one access transistor. To read, the bitline is precharged to VDD/2 and the wordline opens. The cell shares its charge with the much larger bitline capacitance, moving it up for a stored 1 or down for a 0. With Cs = 25 fF, a 150 fF bitline and VDD = 1.8 V, the swing is only 128.6 mV, so a sense amplifier must resolve it. Sharing empties the cell, so every read is followed by a write-back.
Why must DRAM be refreshed, and how often?
Charge leaks from the storage node through the junction and the off access transistor, so a stored 1 sags until the read swing falls below what the sense amplifier can see. With a constant leakage I, the retention time is t = Cs ΔVallowed / I. For the cell above, a 50 mV sense threshold and an illustrative 2 fA leakage give 6.88 s. Leakage rises with temperature, so standard parts (DDR2, for example) refresh every row within 64 ms. With 8,192 rows that is one row every 7.8 µs.
How do the 3T DRAM cell and a ROM differ from the 1T1C cell?
The 3T cell keeps its bit on the gate of a storage transistor. A write transistor puts charge on that gate, and a separate read transistor lets the stored level switch a read bitline. Reading leaves the charge alone, so no write-back is needed, but three transistors cost area. Early parts such as the Intel 1103 used it. A ROM stores bits in the wiring itself. In a NOR ROM, each bitline is precharged high and a transistor is placed only where a 0 is wanted, pulling the bitline low when its word is selected.
How does flash memory store bits, and what are NAND, NOR, SLC and MLC?
A flash cell is a transistor with an extra charge-storing gate. Programming pushes electrons onto it, which raises the threshold voltage. Reading asks whether the cell conducts at a reference gate voltage. NOR flash puts cells in parallel on the bitline for fast random reads. NAND flash chains them in series, which is denser and suits storage. SLC stores 1 bit as 2 threshold levels. MLC stores 2 bits as 4 levels and needs 3 references. In an illustrative 5 V window with ±0.45 V spreads, the gap between levels shrinks from 4.1 V to 0.77 V.
A DRAM's bitline capacitance doubles from 150 fF to 300 fF while the cell stays at 25 fF (VDD = 1.8 V). What happens to the read swing?
ΔV = (VDD/2)·Cs/(Cs + CBL): 0.9 × 25/175 = 128.6 mV becomes 0.9 × 25/325 = 69.2 mV. It falls by a factor 325/175, not quite 2, because Cs is part of the total. That is why bitlines are kept short, with only a few hundred cells each.
Worked examples
Read swing of a 1T1C cell · Cs = 25 fF, CBL = 150 fF, VDD = 1.8 V, bitline precharged to 0.9 V, sense amplifier threshold 50 mV (all illustrative)
- Charge is shared, not lost: Cs·Vcell + CBL·0.9 V = (Cs + CBL)·Vfinal
- Stored 1 (1.8 V): Vfinal = (25 × 1.8 + 150 × 0.9)/175 = 1.0286 V, a swing of +128.6 mV
- Stored 0 (0 V): Vfinal = (150 × 0.9)/175 = 0.7714 V, a swing of −128.6 mV
- Transfer ratio Cs/(Cs + CBL) = 0.143: only one seventh of the cell's half-swing reaches the bitline
- 128.6 mV > 50 mV, so the sense amplifier reads it; an Euler integration of the access resistor (10 kΩ) settles to the same value
Answer: ±128.6 mV on the bitline; readable with margin, and the cell must be rewritten afterwards
Retention time and refresh · The same cell, sense threshold 50 mV, illustrative leakage 2 fA at room temperature and 100 times more when hot; 8,192 rows
- Lowest readable stored 1: the swing (V − 0.9) × 25/175 must stay ≥ 50 mV, so V ≥ 0.9 + 0.05 × 7 = 1.25 V
- Allowed sag: 1.8 − 1.25 = 0.55 V, i.e. 13.75 fC of charge
- At 2 fA: t = 25 fF × 0.55 V / 2 fA = 6.875 s; stepping the leaking cell in 1 ms steps gives the same
- At 200 fA (hot): t = 68.75 ms, close to the 64 ms standard interval
- Refreshing 8,192 rows within 64 ms means one row every 7.8125 µs
Answer: 6.88 s cool, 68.8 ms hot; the 64 ms refresh covers the hot, leaky case
Threshold windows of SLC, MLC and TLC flash · Illustrative: level centres spread evenly over −2 V to +3 V, each level spread ±3σ with σ = 0.15 V
- SLC, 2 levels: spacing 5 V, gap 5 − 0.9 = 4.1 V, one read reference
- MLC, 4 levels: spacing 5/3 = 1.667 V, gap 0.767 V, three references
- TLC, 8 levels: spacing 5/7 = 0.714 V, gap −0.186 V: the levels overlap
- TLC needs σ below (5/7)/6 = 0.119 V just to separate, so programming must be tighter (program and verify in small steps)
- Every extra bit doubles the levels and roughly halves the gap, so reads need finer references and error correction
Answer: gaps 4.1 V (SLC), 0.77 V (MLC), and overlap for TLC unless σ < 0.12 V
Programming a small NOR ROM · 4 words × 4 bits storing 1001, 0011, 1110, 0110 (bit 3 to bit 0); bitlines precharged high
- Selecting a word turns on its row's transistors; any transistor on a bitline pulls it to 0
- So a transistor goes at each crosspoint where the stored bit is 0, and none where it is 1
- Zeros per word: 2, 2, 1, 2
- Transistors in the array: 7 of 16 crosspoints
Answer: 7 transistors; changing the data means changing the mask (or the via layer) that places them
Programmable logic: ROM, PLA, PAL and FPGA look-up tables
Which syllabus covers thisAU EC3552 Unit IVVTU not in syllabusSPPU 404182 Unit IIIMU not in syllabusKTU not in syllabus
How do a ROM, a PLA and a PAL implement logic?
All three are two-level AND-OR arrays. They differ in which plane you can program. A ROM has a fixed AND plane, a full decoder that makes every minterm, and a programmable OR plane, so a 3-input, 2-output function always takes 8 words of 2 bits = 16 bits. A PLA programs both planes, so it builds only the product terms it needs and can share a term between outputs: a full adder needs 7 terms and 18 of the 42 AND crosspoints. A PAL programs the AND plane, but each output ORs its own fixed group of terms: simpler and faster, with no sharing.
How does an FPGA look-up table compute any function of its inputs?
A K-input look-up table (LUT) is a tiny memory: 2^K configuration bits held in SRAM cells, read through a tree of 2^K − 1 two-to-one multiplexers in K levels. The inputs are the address. Level one picks between neighbouring bits using input A, level two uses B, and so on, until exactly one bit reaches the output. Write a function's truth table into the bits and the LUT becomes that function. A 4-LUT has 16 bits and 15 muxes, and can be any of 65,536 four-input functions. A typical FPGA logic cell pairs a 4-input LUT with a full adder and a D flip-flop.
What if a function has more inputs than one LUT?
Split it with Shannon's expansion: f = x'·f(x = 0) + x·f(x = 1). Each half is a function of one input fewer. Joining them is a 2:1 mux, which fits in any LUT with 3 or more inputs. With 4-LUTs, any 5-input function takes at most 3 LUTs (two halves and a mux). Any 6-input function takes at most 7: four quarter-functions and three muxes. In general the bound is 2^(n − K + 1) − 1. Real functions often need far fewer. A 5-input XOR is one 4-LUT for four inputs plus a second LUT that XORs in the fifth.
CPLD or FPGA: what is the difference?
A CPLD is several PAL-like blocks joined by a programmable switch matrix. Their macrocells build sum-of-products logic, and the configuration is non-volatile, so a CPLD works the moment power comes up and can even load an FPGA's configuration. It holds thousands to tens of thousands of gates. A signal crosses one AND-OR array, so its pin-to-pin delay barely depends on placement. An FPGA is a sea of LUT-based logic blocks in a routing mesh. Most keep their configuration in SRAM, reloaded from external flash at every power-up, and they range from tens of thousands to millions of gates. Xilinx's first FPGA, the XC2064 of 1985, had 64 logic blocks.
Shannon's bound says any 5-input function fits in three 4-input LUTs. How many 4-LUTs does a 5-input XOR (parity) need?
XOR is associative: one LUT computes the XOR of the first four inputs, and a second LUT XORs that result with the fifth input. That is two LUTs, one fewer than the general bound, which assumes nothing about the function. A single 4-LUT cannot do it because it has only four inputs.
Worked examples
Program a 3-input LUT as a majority gate (the full adder's carry) · Inputs A, B, C with address = A + 2B + 4C; output 1 when at least two inputs are 1
- Addresses with two or more 1s: 3 (A, B), 5 (A, C), 6 (B, C) and 7 (all three)
- SRAM bits from address 0 to 7: 00010111
- Read A = 1, B = 0, C = 1: address 1 + 0 + 4 = 5
- Level 1 (A = 1) passes the odd bits 1, 3, 5, 7: 0 1 1 1
- Level 2 (B = 0) passes the first of each pair: 0 1; level 3 (C = 1) passes the second: 1
- All 8 input combinations were checked against the majority rule
Answer: Bits 00010111; input A, B, C = 1, 0, 1 reads bit 5, which is 1
A full adder in a ROM and in a PLA · Outputs S = A ⊕ B ⊕ C and Co = majority(A, B, C)
- ROM: a fixed 3-to-8 decoder and 8 words of 2 bits: 16 programmed bits, whatever the function
- PLA sum: S is 1 for minterms A'B'C, A'BC', AB'C' and ABC: 4 product terms of 3 literals
- PLA carry: Co = AB + AC + BC: 3 product terms of 2 literals
- AND plane: 6 literal columns (A, A', B, B', C, C') × 7 terms = 42 crosspoints, 18 used (4 × 3 + 3 × 2)
- OR plane: 2 outputs × 7 terms = 14 crosspoints, 7 used; all 8 inputs give S and Co equal to the binary sum A + B + C
Answer: ROM 16 bits; PLA 7 product terms with 18 of 42 AND and 7 of 14 OR crosspoints
A 6-input function from 4-input LUTs · Any function f of inputs a to f; 4-input LUTs only
- Expand on e and f: f = e'f'·f00 + e f'·f10 + e'f·f01 + e f·f11, four cofactors of a, b, c, d only
- Each cofactor is one 4-LUT: 4 LUTs
- Two 2:1 muxes switched by e join them in pairs, and one mux switched by f joins the pair: 3 LUTs used as muxes
- Total 7 LUTs = 2^(6 − 4 + 1) − 1, holding 7 × 16 = 112 configuration bits; a single 6-input LUT would need 64
- The network was checked against the truth table on all 64 inputs for 20 random functions
Answer: 7 four-input LUTs, two levels of muxes after the cofactor LUTs
Verilog HDL, test benches and RTL design of a state machine
Which syllabus covers thisAU EC3552 Unit VVTU not in syllabusSPPU 404182 Unit IMU ECC503 Module 6KTU PEECT415 Modules 1, 4
What is in a Verilog module, and what are wire and reg?
A module is a block of hardware with named ports: input, output and inout, each one bit wide or a vector such as [5:0]. Inside, a wire is a connection with no memory. It always shows whatever drives it, a gate output or an assign statement. A reg holds the value an always or initial block last gave it, and it becomes a flip-flop only when it is assigned on a clock edge. localparam names constants such as state codes. Verilog was standardised as IEEE 1364 (1995, 2001, 2005), then merged into SystemVerilog, IEEE 1800, in 2009.
What are gate-level, dataflow and behavioural modelling?
They are three ways to describe the same hardware. A 2:1 mux at gate level is built from primitives: a not, two ands and an or. In dataflow style it is one continuous assignment, assign y = s ? b : a;. Behaviourally it is a procedure, always @(*) if (s) y = b; else y = a;. Synthesis turns all three into the same gates. In clocked blocks use non-blocking <=: every right-hand side is read before any register changes, just as real flip-flops all sample on the same clock edge. Blocking = updates at once and belongs in combinational blocks.
How do you design a coin-operated dispenser as a controller plus a datapath?
RTL design splits the job in two. The datapath holds the data: a 6-bit register tot for the money so far, a coin register cv, an adder and a comparator for tot ≥ s. The controller is a state machine that tells the datapath what to do on each clock. Init clears tot, Wait takes a coin or spots tot ≥ s, Add loads tot + cv, and Disp raises d for one cycle with chg = tot − s. Set a ₹25 price and feed in three ₹10 coins: the machine dispenses on clock edge 8 with ₹5 change. The total never passes s + 9 = ₹39, so 6 bits are enough.
What is a test bench, and how does the same register look in VHDL?
A test bench is a Verilog module with no ports. It is only simulated, never built. It instantiates the design and makes a clock with always #5 clk = ~clk;, a 10 ns period with a 1 ns time unit. An initial block drives the inputs at chosen times, and the bench compares outputs with expected values, printing a message on any mismatch. KTU's syllabus uses VHDL (IEEE 1076, first published in 1987) instead. There the register always @(posedge clk) q <= d; becomes process (clk) begin if rising_edge(clk) then q <= d; end if; end process; inside an architecture.
Inside always @(posedge clk) you write a <= b; b <= a;. Before the edge a = 1 and b = 0. What are a and b after the edge?
Non-blocking assignments read every right-hand side first (b = 0, a = 1), then update all the registers together, like two flip-flops clocked by the same edge. So a takes 0 and b takes 1. With blocking = the first line would change a before the second line read it, and both would end up 0.
Worked examples
Trace the soda machine through one sale · Price s = ₹25; three ₹10 coins wait at the sensor (each one is taken in a Wait state); the machine starts from reset in Init
- Edge 1: Init → Wait, tot = 0
- Edge 2: Wait sees c = 1 and tot = 0 < 25, so it latches cv = 10 and goes to Add
- Edge 3: Add loads tot = 0 + 10 = 10 and returns to Wait; edges 4-5 repeat this (tot = 20)
- Edges 6-7: the third coin, tot = 30
- Edge 8: Wait finds tot = 30 ≥ 25 and goes to Disp, so d = 1 and chg = 30 − 25 = ₹5
- Edge 9: Disp → Init, which clears tot on the next edge
- Checked: a Python and a TypeScript model both follow the rule 'add coins until the total reaches the price, then dispense with the excess as change' on every sequence of up to 6 coins at every price from ₹15 to ₹30
Answer: Dispenses on edge 8 with ₹5 change
A different price, and how many flip-flops · Price s = ₹18; coins ₹5, ₹5, ₹10; registers st, tot and cv as in the Verilog
- Totals after each Add: 5, 10, 20
- 20 ≥ 18 in the next Wait: Disp on edge 8, chg = 20 − 18 = ₹2
- Largest possible total: the last coin arrives with tot ≤ s − 1 and adds at most 10, so tot ≤ s + 9; at s = 30 that is ₹39, which needs 6 bits (up to 63)
- Four states need ⌈log2 4⌉ = 2 state bits; cv is 5 bits to match the coin input a
- Flip-flops: 2 + 6 + 5 = 13
Answer: Change ₹2 on edge 8; the design uses 13 flip-flops
When does d rise in simulation? · Test bench clock always #5 clk = ~clk;, clk starting at 0, time unit 1 ns; the sale of the first example
- The clock toggles every 5 ns: period 10 ns, frequency 100 MHz
- Rising edges fall at 5, 15, 25 ns and so on: edge k at 10k − 5 ns
- The machine enters Disp on edge 8: t = 10 × 8 − 5 = 75 ns
- d stays high for one clock period, until the edge at 85 ns
Answer: d is high from 75 ns to 85 ns
The ASIC design flow, wafer to chip, dies per wafer and yield
Which syllabus covers thisAU EC3552 Unit VVTU BEC602 Modules 3, 5SPPU 404182 Unit VMU ECC503 Module 1KTU PEECT415 Module 4
What happens between RTL and a design ready for tape-out?
The ASIC flow is a chain of tools, each one checked before the next. RTL is simulated against a test bench. Synthesis maps it onto a standard-cell library under timing constraints, giving a gate netlist. Physical design floorplans the die, places the cells, builds the clock tree and routes the wires. Extraction turns the real wires into resistances and capacitances, and static timing analysis uses them to check setup and hold on every path. Physical verification runs DRC (does the layout obey the foundry's rules?) and LVS (does it match the netlist?). Only then is the layout taped out, usually as a GDSII file, and sent off to make the masks.
How does a processed wafer become packaged chips?
The fab builds every die on the wafer at once, layer by layer, through repeated lithography, etching, implantation and deposition. Before the wafer is cut, wafer probe tests each die through needles on its pads and marks the failures. A saw then dices the wafer along the scribe lines between dies. Good dies are attached to packages and connected by bond wires or by solder bumps (flip-chip), and each packaged part is tested again. Wafers grew so that each run makes more dies: 150 mm wafers arrived in 1983, 200 mm in 1992 and 300 mm in 1999.
How many dies fit on a wafer?
Divide the wafer area by the die area, then subtract the partial dies lost round the edge, roughly one for every die diagonal of circumference. With d the usable diameter and A the die area, gross dies ≈ πd²/4A − πd/√(2A). Take a 300 mm wafer with a 3 mm edge exclusion (illustrative) and 10 × 10 mm dies: the formula gives 613, and counting a real grid gives 621. The formula is slightly pessimistic, and small dies waste less of the edge. The interactive counts the grid exactly and shows which dies a random scatter of defects kills.
How do defects set the yield, and why does die size matter so much?
Suppose point defects land at random with density D0. A die of area A then gets AD0 defects on average and escapes them all with probability e^(−AD0): the Poisson model. Real defects cluster, which leaves more dies untouched. Murphy's model averages over a spread of densities, and the negative binomial adds a clustering factor α. At D0 = 0.5 per cm², a 1 cm² die yields 60.7 % (Poisson), 61.9 % (Murphy) or 63 % (α = 3). Double the area and the Poisson yield drops to 36.8 %, so big dies lose twice: fewer per wafer, and a smaller share of them good.
A 300 mm wafer carries 10 × 10 mm dies at D0 = 0.5 per cm² (Poisson). The design grows to 10 × 20 mm. Roughly what happens to the number of good dies per wafer?
Gross dies fall from 621 to 300 (just under half), and yield falls from e^(−0.5) = 60.7 % to e^(−1) = 36.8 %. Good dies go from about 377 to about 110, 29 % of before. Doubling the area costs twice: fewer dies, and each one is more likely to catch a defect.
Worked examples
Dies on a 200 mm wafer · Wafer 200 mm, edge exclusion 3 mm (usable diameter 194 mm), die 8 × 6 mm (A = 48 mm²)
- Area term: π × 194² / (4 × 48) = 615.8
- Edge term: π × 194 / √(2 × 48) = 62.2
- Formula: 553.6, about 553 dies
- Exact grid count (best of four grid placements, every die fully inside 97 mm): 568
- The formula is 2.5 % low: it assumes an average edge loss, and a well-placed grid does better
Answer: About 553 by the formula, 568 by exact count
Yield of that die by three models · Die area A = 0.48 cm², D0 = 0.8 defects per cm², 568 gross dies, clustering α = 3
- AD0 = 0.48 × 0.8 = 0.384 defects per die on average
- Poisson: e^(−0.384) = 68.11 %
- Murphy: ((1 − e^(−0.384)) / 0.384)² = 68.95 %
- Negative binomial: (1 + 0.384/3)^(−3) = 69.67 %
- Good dies: 387, 392 and 396
- A Monte Carlo of 300 wafers with random defects gives 68.34 % good, matching Poisson
Answer: 68.1 % (Poisson), 69 % (Murphy), 69.7 % (α = 3): about 387 to 396 good dies
Cost per good die · 300 mm wafer costing USD 6,000 (illustrative), 3 mm edge exclusion, die 12 × 12 mm, D0 = 0.3 per cm², negative binomial with α = 3
- Gross dies: formula 417, exact grid 432
- A = 1.44 cm², AD0 = 0.432; yield (1 + 0.432/3)^(−3) = 66.79 %
- Good dies: 432 × 0.6679 = 288.5
- Cost per good die: 6,000 / 288.5 = USD 20.79
- With Poisson (64.92 %) the estimate would be USD 21.39; packaging and test add to both
Answer: About 289 good dies, USD 20.79 each before packaging and test
Testing: stuck-at faults, ATPG, scan design, BIST and boundary scan
Which syllabus covers thisAU EC3552 Unit VVTU BEC602 Module 5SPPU 404182 Unit VIMU not in syllabusKTU PEECT415 Module 3
What is a stuck-at fault, and how does one test vector catch it?
The stuck-at model treats a manufacturing defect as one line tied permanently to 0 (s-a-0) or to 1 (s-a-1). A vector catches the fault only if it does two things. It must activate the fault, driving the line to the opposite value. Then it must propagate the difference to an output, holding the other inputs of every gate on the way at their non-controlling value (1 for a NAND). D marks a line that is 1 in the good circuit and 0 in the faulty one, and D' the reverse. The c17 benchmark has 17 lines (11 nets plus 6 fanout branches), so it has 34 single stuck-at faults.
How are test vectors found, and what is fault coverage?
Fault coverage is the share of modelled faults that a test set detects. One way to find tests is the Boolean difference, ∂f/∂x = f(x = 0) ⊕ f(x = 1), which is 1 exactly when the output depends on x. So x s-a-0 is detected by the vectors with x = 1 and ∂f/∂x = 1. ATPG tools reach the same answer by sensitising paths. Equivalent faults need only one test: a NAND input s-a-0 looks just like its output s-a-1, which collapses c17's 34 faults into 22 classes. Nor do you need all 32 input vectors: 4 well-chosen vectors detect all 34 faults.
Why do flip-flops make testing hard, and how does scan design fix it?
Inside a sequential circuit you cannot set or read the flip-flops directly, so reaching a fault can take a long input sequence. Scan design adds a 2:1 mux at every flip-flop's D input. With scan enable on, the flip-flops form one long shift register. You shift a chosen state in, pulse one normal clock to capture the logic's response, and shift it out while the next pattern shifts in. The logic between flip-flops can then be tested as if it were combinational. The price is a mux per flip-flop, plus test time. With 1,000 flip-flops in one chain, 200 patterns take 201,200 clocks; ten chains of 100 take 20,300.
How does a chip test itself (BIST)?
Built-in self-test puts the tester on the chip. A linear-feedback shift register (LFSR) generates the patterns. It is a shift register whose input is the XOR of a few of its own bits. With a primitive feedback polynomial it visits every state except all-zeros before repeating, so the 4-bit LFSR for x⁴ + x³ + 1 runs through 15 states. A multiple-input signature register, also an LFSR, compresses the responses into a short signature, which is compared with a good chip's signature. On c17, a 5-bit LFSR (x⁵ + x³ + 1, 31 states) reaches 100 % fault coverage after 14 patterns.
What is boundary scan (JTAG, IEEE 1149.1)?
IEEE 1149.1-1990, the Standard Test Access Port and Boundary-Scan Architecture, puts a shift-register cell at every signal pin. Chained together, these cells let a tester set and read the pins of every chip on a board without probes, which finds opens and shorts in the solder joints. Four pins form the test access port: TDI (data in), TDO (data out), TCK (clock) and TMS (mode select), plus an optional TRST reset. TMS steps a 16-state TAP controller, and five clocks with TMS high always reach Test-Logic-Reset. BYPASS, SAMPLE/PRELOAD and EXTEST are mandatory instructions.
c17 has 5 inputs, so exhaustive testing applies 32 vectors. What is the smallest number of vectors that detects all 34 single stuck-at faults?
Each vector detects several faults at once (between 4 and 14 of the 34 here). An exhaustive search finds that no 3 vectors cover every fault, but 4 do, for example 10000, 01010, 00101, 11111 (inputs in the order 1 2 3 6 7). Large circuits rely on the same effect: ATPG test sets are far smaller than 2^n.
Worked examples
Find a test for a fanout-branch fault · c17, branch 11→19 (net 11 feeding gate 19) stuck at 1
- Activate: the good value on the branch must be 0, so net 11 = NAND(3, 6) = 0, which needs inputs 3 = 1 and 6 = 1
- Propagate through gate 19 = NAND(11, 7): set 7 = 1 (non-controlling); good 19 = NAND(0, 1) = 1, faulty NAND(1, 1) = 0, so line 19 carries D
- Propagate through gate 23 = NAND(16, 19): net 16 = NAND(2, 11) is 1 because the branch into gate 16 still sees 11 = 0, so 23 = NAND(1, D) = D'
- Inputs 1 and 2 are free
- Fault simulation confirms exactly 4 vectors detect it: 00111, 10111, 01111, 11111 (order 1 2 3 6 7)
Answer: Any vector with inputs 3 = 6 = 7 = 1; output 23 reads 1 instead of 0
Boolean difference with respect to input 3 · c17 outputs: f22 = x1·x3 + x2·(x3·x6)' and f23 = (x3·x6)'·(x2 + x7)
- f22 with x3 = 0: x2. With x3 = 1: x1 + x2·x6'
- ∂f22/∂x3 = x2 ⊕ (x1 + x2·x6') = x1·x2' + x1'·x2·x6
- f23 with x3 = 0: x2 + x7. With x3 = 1: x6'·(x2 + x7). So ∂f23/∂x3 = (x2 + x7)·x6
- Input 3 s-a-0 needs x3 = 1 and either difference equal to 1
- Counting over all 32 vectors: 9 detect it, the same set that serial fault simulation finds
Answer: ∂f22/∂x3 = x1·x2' + x1'·x2·x6, ∂f23/∂x3 = x6·(x2 + x7); 9 vectors detect input 3 s-a-0
The states of a 4-bit LFSR · Feedback polynomial x⁴ + x³ + 1: shift right, new top bit = bit 0 ⊕ bit 1; seed 0001 (bits b3 b2 b1 b0)
- States 1-8: 0001, 1000, 0100, 0010, 1001, 1100, 0110, 1011
- States 9-15: 0101, 1010, 1101, 1110, 1111, 0111, 0011
- The next state is 0001 again: period 15 = 2⁴ − 1, and all 15 non-zero states appear once
- All-zeros maps to itself (the XOR of zeros is 0), which is why the seed must not be 0000
- The non-primitive x⁴ + x² + 1 (new top bit = bit 0 ⊕ bit 2) repeats after only 6 states from 0001: not every polynomial gives maximal length
Answer: A maximal-length sequence of 15 states from 0001 back to 0001
Keep going
- Digital Electronics interactive book
- Microcontrollers interactive book
- Signals and Systems interactive book
- Pre-Graduate Program in Embedded Systems
- Analog Electronics interactive book
- All Skill-Lync interactive books
Recommended reading
- Digital Integrated Circuits: A Design Perspective, J. M. Rabaey, A. Chandrakasan and B. Nikolic
- CMOS VLSI Design: A Circuits and Systems Perspective, N. H. E. Weste and D. M. Harris
- CMOS Digital Integrated Circuits: Analysis and Design, S.-M. Kang and Y. Leblebici
- Principles of CMOS VLSI Design, N. H. E. Weste and K. Eshraghian
- Digital Design with RTL Design, VHDL and Verilog, F. Vahid
- Essentials of Electronic Testing, M. L. Bushnell and V. D. Agrawal
How this book was made
Chapters follow the official syllabi listed above. Every interactive's formulas and every worked-example number were checked by script. Explanations, animations and examples are original to Skill-Lync. Topics informed by MIT OpenCourseWare 6.374 (Analysis and Design of Digital Integrated Circuits), CC BY-NC-SA. All explanations, animations and examples are original to Skill-Lync; MIT does not endorse this page.