VLSI Design Techniques: Design Styles, Low-Power Methods and DFT
VLSI design techniques decide a chip's power, performance and area. This guide compares design styles (full-custom, standard-cell, gate array, FPGA), explains clock gating, multi-Vt, power gating, DVFS, pipelining and DFT, and works one block's power from 517 mW down to 135 mW.

In short
- VLSI design techniques are the choices that decide a chip's power, performance and area (PPA): the design style (full-custom, standard-cell, gate array, FPGA), how logic is synthesised and laid out, how power is cut, and how the chip is made testable.
- Switching power follows . Voltage enters squared, so lowering the supply is the strongest single lever; clock gating attacks , the activity.
- Style matters as much as technique: averaged over benchmark circuits, an FPGA needed about 35× the area of a standard-cell ASIC, ran roughly 3 to 4× slower and used about 14× the dynamic power for core logic (Kuon and Rose, IEEE TCAD, Feb 2007).
- In Skill-Lync's worked example below, clock gating, multi-Vt cells and a DVFS light-load point take one block from 517 mW to 135 mW.
VLSI design techniques are the methods engineers use to turn a logic idea into a chip that is small, fast enough and frugal with power. They cover four decisions: which design style to build in (full-custom, standard-cell, gate array or FPGA), how the RTL is synthesised and laid out, which low-power methods are applied, and how the finished silicon will be tested. Getting them right is the difference between a phone chip that lasts a day and one that overheats.
This guide explains each technique in plain terms, shows where it enters the design flow, and works through one block's power budget with real formulas so you can see what each method saves. It is written for B.Tech ECE and EEE students and freshers preparing for VLSI interviews. Skill-Lync computed every number in the example with a short Python script.
What are VLSI design techniques?
VLSI design techniques are the methods that decide how a very-large-scale integrated circuit meets its targets for power, performance and area, the trio engineers call PPA. They include the design style, logic synthesis, floorplanning and layout, low-power methods such as clock gating and voltage scaling, timing optimisation, and design-for-test. Each one trades one PPA goal against another.
VLSI (very-large-scale integration) means placing a very large number of MOS transistors on one die, a step that followed small-, medium- and large-scale integration in the 1970s. No engineer draws billions of transistors by hand, so the techniques are as much about method as circuits: describe the design at a higher level, let tools map it to cells, and reserve hand work for the few blocks where it pays. If the CMOS basics are new to you, the CMOS logic family chapter of Skill-Lync's free Digital Electronics book shows how one NAND gate is built from four transistors.
Which VLSI design styles are there, and how do they compare?
There are four main styles. Full-custom design draws every transistor for the best speed and density. Standard-cell (semi-custom) design builds the chip from a library of pre-designed gates. Gate arrays customise only the metal layers over prefabricated transistors. FPGAs are bought ready-made and configured in the field. Moving from full-custom to FPGA trades efficiency for lower cost and faster turnaround.
| Style | What is custom | Strength | Cost of that strength | Typical use |
|---|---|---|---|---|
| Full-custom | Every transistor and wire | Highest speed and density | Longest design time, highest engineering cost | CPU datapaths, SRAM bit cells, analog and I/O blocks |
| Standard-cell (semi-custom) | Placement and wiring of library cells | Good PPA with automated tools | Full mask set, months of verification | Most ASICs and SoCs |
| Gate array / structured ASIC | Only the top metal layers | Fewer masks, faster turnaround | Wasted transistors, lower density | Medium-volume parts |
| FPGA (programmable) | Nothing in silicon: a configuration file | No fabrication, can be changed after shipping | Area, speed and power penalty (see below) | Prototypes, low volume, changing standards |
The FPGA penalty has been measured. Ian Kuon and Jonathan Rose compared the same benchmark circuits in a 90 nm FPGA and a 90 nm standard-cell ASIC. For circuits built only from look-up-table logic and flip-flops, the FPGA needed about 35 times the area. Its critical paths were roughly three to four times slower, and it used about 14 times the dynamic power. Hard multiplier and memory blocks narrowed the area gap to as little as 18 times (IEEE TCAD, vol. 26, Feb 2007). That is why FPGAs prototype a design and ASICs ship it in volume. The CPLD and FPGA chapter explains how the configurable logic blocks work.
Where do these techniques enter the design flow?
Each technique belongs to a stage. Architecture fixes pipelining, parallelism and power domains. RTL coding adds clock enables. Synthesis maps logic to cells, inserts clock-gating cells and scan chains, and picks threshold voltages. Place and route adds power switches and builds the clock tree. Signoff checks timing, IR drop and test coverage. Earlier choices usually save more.
For the flow itself, from RTL to the GDSII file sent to the foundry, read the steps involved in VLSI design. How a block is checked before layout is covered in circuit simulation and analysis in VLSI design.
Logic synthesis
Logic synthesis turns RTL code (Verilog or VHDL) into a gate-level netlist built from a standard-cell library: NAND, NOR, flip-flops, multiplexers, adders. The tool optimises against constraints you give it, chiefly the clock period, and trades area for speed where a path is too slow. Writing RTL that synthesises well is a skill in itself; the Verilog chapter of the Digital Electronics book covers the coding styles synthesis tools expect.
Floorplanning, placement and routing
Floorplanning decides where each large block sits and where power rails run. Placement fixes every cell's position, and routing draws the metal wires between them. Good floorplans keep blocks that talk often close together, which shortens wires. Shorter wires mean less capacitance, which means both less delay and less power.
How do chip designers reduce power consumption?
Chip power has two parts. Dynamic power is spent charging and discharging capacitance each time a node switches. Static power is leakage that flows even when nothing switches. Designers cut dynamic power by lowering voltage, frequency or switching activity, and cut leakage by using slower, high-threshold transistors or switching idle blocks off completely.
The standard model for switching power is
where is the activity factor (the fraction of the capacitance that switches each cycle), the switched capacitance, the supply voltage and the clock frequency. Total power is . Leakage has grown in importance as transistors shrank. Every technique below attacks one term of these equations.
Clock gating: stop the clock where nothing changes
The clock net switches every cycle, so its activity factor is 1, the highest on the chip. One vendor estimate puts the clock network at more than half of a chip's power (Mentor Graphics, 2012). Clock gating inserts an integrated clock-gating (ICG) cell, a latch plus an AND gate, in front of a group of registers. When their enable is low, the clock stops and those registers, and the logic they drive, stop switching. Synthesis tools insert ICGs automatically from if (enable) patterns in RTL.
Multi-threshold (multi-Vt) cells: fast where needed, low-leakage elsewhere
A transistor's threshold voltage sets a trade-off. Low-Vt cells switch faster but leak much more; high-Vt cells are slower but typically leak about ten times less. Tools place low-Vt cells only on critical timing paths and high-Vt cells everywhere else, so most of the chip leaks little without missing the clock target.
Power gating: switch idle blocks off
Power gating puts header (PMOS) or footer (NMOS) sleep transistors between a block and its supply. When the block is idle they open and its leakage nearly vanishes. The price: wake-up time, the area of the switches, isolation cells so that floating outputs don't corrupt powered-on neighbours, and retention flip-flops if the block must remember its state.
Dynamic voltage and frequency scaling (DVFS)
DVFS lowers the supply voltage and clock together when the workload is light. Lower voltage makes gates slower, so frequency has to drop with it. Because power goes with , a modest voltage drop plus a frequency cut saves a lot. Mobile processors switch between several characterised voltage-frequency operating points many times a second.
Pipelining and parallelism: buy speed, then trade it for voltage
Pipelining splits a long logic path with registers so each stage finishes sooner; parallelism duplicates hardware so each copy can run slower. Either way, the design meets its throughput at a lower clock or with spare timing slack, and that slack is spent on a lower supply voltage. Chandrakasan, Sheng and Brodersen showed in 1992 that this architecture-driven voltage scaling can improve energy efficiency by about an order of magnitude, at the cost of extra area (IEEE JSSC, April 1992).
How is power intent described to the tools?
Power domains, switches, isolation and retention are not written in the RTL. They go in a separate power-intent file in the Unified Power Format, standardised as IEEE 1801. The current edition, IEEE 1801-2024 (UPF 4.0), is free to download through the IEEE GET Program, announced by Accellera in March 2025. Simulation, synthesis and place-and-route all read the same UPF file, so the power plan stays consistent.
Worked example: one block's power, technique by technique
Take a digital block running at 0.90 V and 1 GHz. Assume its clock network switches 0.30 nF every cycle () and its logic and registers have 2.0 nF of capacitance at 12% activity. With every cell low-Vt, leakage is 80 mW. These inputs are illustrative but in a realistic range; the method is what matters.
Dynamic power is 437.4 mW, 55.6% of it in the clock, and the total with leakage is 517.4 mW. Now apply the techniques one at a time:
- Clock gating. Suppose 70% of the clock capacitance sits behind ICG cells and those units are idle half the time. The clock term falls by a factor of , to 158.0 mW. That saves 85.0 mW, 19.4% of dynamic power. (We left logic power unchanged, which is conservative: gated registers also stop driving their logic.)
- Multi-Vt. Move 80% of cells to high-Vt, at about one-tenth the leakage. Leakage falls from 80 mW to mW. The block now draws 374.8 mW, 28% less than the start, at the same speed.
- DVFS at light load. Drop to an operating point of 0.72 V and 500 MHz. Dynamic power scales by , to 112.8 mW. With leakage held at 22.4 mW, the block draws about 135 mW.
Here is the catch in step 3. Halving the frequency alone, at 0.90 V, halves power but not the energy per task: the same work takes twice as long. Only the voltage term lowers the energy of each operation, here by , a 36% saving per task. Leakage makes a slow-but-full-voltage mode worse still, because the chip leaks for longer. That is why DVFS lowers V and f together, and why "race to idle, then power-gate" sometimes beats running slowly.
We held leakage constant at the lower voltage, which overstates it a little; real leakage also falls with voltage and rises steeply with temperature. Power gating would remove the leakage of any block switched off: a block holding a quarter of the leakage in this example would save another 5.6 mW while it sleeps.
How do designers meet timing efficiently?
Designers meet timing by shortening the slowest register-to-register paths. They pipeline long paths, resize gates on critical paths, use low-Vt cells there, restructure logic, and build a clock tree with little skew. Static timing analysis then checks every path at every process, voltage and temperature corner for setup and hold violations, without needing simulation vectors.
The trade-off is always PPA. Upsizing a gate cuts its delay but adds area and input capacitance, which costs power. Low-Vt cells are fast but leak. A deeper pipeline raises throughput but adds latency and more registers, which means more clock load. The flip-flop timing section of the Digital Electronics book covers setup and hold time, the two constraints every path must meet.
What is design for testability (DFT) in VLSI?
Design for testability adds hardware that makes a manufactured chip easy to test. Scan design links the flip-flops into shift registers called scan chains, so a tester can load any internal state, clock the logic once and shift the result out. Automatic test pattern generation (ATPG) creates the patterns. Built-in self-test (BIST) lets memories and logic test themselves.
Without DFT, most gates sit deep inside the chip, far from any pin, and reaching a given internal state would need long, hard-to-find input sequences. Scan turns that sequential problem into a much easier combinational one. The cost is a multiplexer in front of every scan flip-flop, a little extra delay and wiring, and test time on the tester. Every production ASIC pays it, because shipping a defective chip costs far more.
Analog vs digital VLSI design: what changes?
Digital VLSI design works with two logic levels and is highly automated: RTL, synthesis, standard cells and automatic place-and-route. Analog VLSI design handles continuously varying voltages and currents, as in amplifiers, data converters, PLLs and voltage regulators. It is largely hand-crafted transistor by transistor, because gain, noise, offset and matching cannot be left to a synthesis tool.
| Digital VLSI | Analog VLSI | |
|---|---|---|
| Signals | Two levels (0 and 1) with noise margins | Continuously varying voltage or current |
| Design entry | RTL code, synthesised to standard cells | Transistor-level schematics, custom layout |
| Main checks | Static timing, functional verification, DFT coverage | SPICE simulation of gain, bandwidth, noise, offset, matching |
| Examples | CPUs, GPUs, DSP datapaths, memory controllers | ADCs, DACs, PLLs, LDO regulators, RF front ends |
Digital is not immune to noise: crosstalk, IR drop and ground bounce can still flip a bit if the noise margin is exceeded. Most chips today are mixed-signal, so a VLSI team needs both kinds of engineer.
How can a student start learning VLSI design techniques?
Start with digital logic: Boolean algebra, combinational and sequential circuits, flip-flop timing and CMOS gates. Then learn Verilog and write synthesisable RTL, followed by static timing and the basics of the flow. Practise with open-source tools and solved GATE questions before applying for verification or physical design roles.
- Digital fundamentals, free: Skill-Lync's interactive Digital Electronics book, then the solved GATE digital electronics questions for practice with worked answers.
- A guided course: Skill-Lync's Basics of Digital Electronics course, the digital-logic foundation for VLSI work.
- Interview preparation: top VLSI interview questions and answers and VLSI engineer career and salary prospects in India.
Not sure whether VLSI, embedded systems or both suit you? Talk to a Skill-Lync counsellor about which course fits your branch and year.
Frequently asked questions
What are the main low-power techniques in VLSI?
The main techniques are clock gating, which stops the clock to idle registers; multi-threshold cells, which use slow, low-leakage transistors off the critical path; power gating, which switches idle blocks off; dynamic voltage and frequency scaling; and multi-voltage domains. At the architecture level, pipelining and parallelism let a design meet throughput at a lower supply voltage.
Why does voltage matter more than frequency for power?
Switching power is , so it scales with the square of voltage but only linearly with frequency. Lowering frequency alone stretches the same work over more time, leaving energy per task unchanged. Lowering voltage cuts energy per operation: going from 0.90 V to 0.72 V saves 36% per operation. Frequency must drop too, because gates slow down.
What is the difference between clock gating and power gating?
Clock gating stops the clock to idle registers, which removes their dynamic switching power. The block stays powered and keeps its state, and it wakes within a cycle. Power gating cuts the supply to an idle block, which also removes leakage. Waking takes longer, and it needs isolation cells, plus retention registers if state must survive.
Is FPGA design part of VLSI?
Yes. An FPGA is itself a VLSI chip, and designing for one uses the same RTL, synthesis and timing analysis skills. The difference is that the FPGA user configures prefabricated logic blocks instead of fabricating new silicon. FPGAs suit prototypes and low volumes, while ASICs win on area, speed and power in high volume.
What is PPA in VLSI?
PPA stands for power, performance and area, the three goals every VLSI design balances. Performance is usually the achievable clock frequency or throughput, power is dynamic plus leakage, and area is silicon size, which drives cost per die. Most techniques improve one at some cost to another, so design is about choosing the right trade-off.
Which skills do VLSI freshers need for these techniques?
Freshers need strong digital logic, Verilog or SystemVerilog, static timing basics (setup, hold, clock skew), CMOS fundamentals, and scripting in Tcl or Python to drive EDA tools. Verification roles add SystemVerilog and UVM. Physical design roles add floorplanning, clock tree synthesis and power analysis. Low-power concepts such as clock gating and UPF come up in interviews for both.
Sources and further reading
- I. Kuon and J. Rose, Measuring the Gap Between FPGAs and ASICs, IEEE Transactions on Computer-Aided Design, vol. 26, no. 2, Feb 2007 (90 nm FPGA vs 90 nm standard-cell ASIC: 35× area, delay ratio roughly 3-4, about 14× dynamic power; accessed 17 Nov 2026)
- A. P. Chandrakasan, S. Sheng and R. W. Brodersen, Low-Power CMOS Digital Design, IEEE Journal of Solid-State Circuits, vol. 27, no. 4, April 1992 (architecture-driven voltage scaling; accessed 17 Nov 2026)
- Accellera, IEEE Standard 1801-2024 (UPF) available through the IEEE GET Program, March 2025 (accessed 17 Nov 2026)
- Wikipedia, Dynamic voltage scaling (switching power C·V²·A·f; accessed 17 Nov 2026)
- Wikipedia, Multi-threshold CMOS (high-Vt devices typically reduce leakage about 10×; accessed 17 Nov 2026)
- Wikipedia, Power gating and Design for testing (accessed 17 Nov 2026)
- A. Narayanan (Mentor Graphics), The trouble with clock trees, Semiconductor Engineering, 11 Oct 2012 (vendor estimate: clock network more than half of chip power; accessed 17 Nov 2026)

