Access Official Technical PDF

Share Summary:LinkedInEmail

ASIC power optimization does not end when RTL is complete. During physical implementation, synthesis, cell selection, placement, clock tree synthesis, routing and sign-off optimization all influence power, performance and area (PPA). For engineers and technical decision-makers, the challenge is to evaluate these choices without sacrificing operating frequency or exceeding area constraints.

Evaluation of several implementation options on an advanced CPU in a 7 nm process illustrates an important principle: a technique’s expected benefit is not a substitute for measurement. Some options can reduce power, while others produce little observable improvement or increase implementation runtime. Results depend on the design, technology libraries and EDA flow.

Why physical implementation matters to ASIC power optimization

ASIC implementation commonly divides into front-end and back-end work. Synthesis converts an RTL description into a gate-level netlist and maps logic to cells from a specific technology library. The back end places those cells, builds the clock distribution network, routes signals and performs further timing and sign-off checks.

Each stage makes decisions that affect PPA. For example, timing requirements influence which cells are selected, while clock tree synthesis inserts buffers to distribute the clock and control skew—the difference in clock arrival time between sequential elements. Added buffers can increase both cell area and power. Later optimization may revise earlier choices, so the complete flow, rather than an isolated setting, should be evaluated.

Distinguish dynamic power from leakage

Power analysis is more useful when its components are considered separately. Dynamic power is associated with circuit switching, including the charging and discharging of capacitive loads. Static or leakage power is consumed while logic is not changing state. Leakage is a particular concern as transistor dimensions scale, and techniques that reduce leakage can have different consequences for speed and dynamic power.

Power estimates also depend on switching activity. Reports based on assumed activity may not represent how a design behaves under a particular workload. In the evaluated methodology, gate-level simulation generated switching activity for power analysis, using a Dhrystone benchmark. This approach considers timing information at the gate level, although it requires more processing time than a zero-delay RTL simulation.

Multibit flip-flops can shrink clock-tree demand

Sequential elements such as D-type flip-flops share the clock network. When individual flip-flops are grouped in suitable configurations, an EDA tool may merge them into multibit flip-flops. The rationale is to reduce the clock-distribution burden: fewer separate clock connections can mean fewer buffers needed to manage clock skew, potentially lowering clock-related power and area.

In the reported trial, enabling two-bit flip-flop merging reduced dynamic power attributed specifically to flip-flops by 25% and total dynamic power by 3.5%. Flip-flop leakage power increased by 24%, corresponding to a 7% rise in total leakage power. Across dynamic and leakage components, total power fell by 2.5%. Standard-cell area improved by 0.6%, and operating frequency remained in the same range.

Approximately 90% of single-bit flip-flops were considered for merging in that design. The result is useful as a case study, not a universal prediction: the balance between clock-related savings and leakage depends on the cells selected and the design’s structure. The trial also found no considerable added implementation runtime from enabling the feature.

Choose threshold-voltage and channel options deliberately

Technology libraries may offer standard cells with different threshold voltages (VT). In general, low-VT cells switch faster but have higher leakage; high-VT cells reduce leakage but have greater delay. Standard-VT cells provide an intermediate option. A design can use these choices to balance timing-critical paths against areas where delay is less restrictive.

Libraries may also offer different channel-length options. The evaluated technology included faster cells with higher leakage and lower dynamic power, and slower cells with lower leakage and slightly higher dynamic power. These trade-offs mean that the “lowest-power” choice is not necessarily consistent across every power component or path.

Trials varied when multi-VT and channel options were made available during synthesis, place and route, and sign-off optimization. PPA differences were small and did not identify one configuration as consistently superior. In one comparison, enabling all cell options earlier increased runtime by 25% during synthesis and 12% during place and route, with a 3% increase at sign-off. A configuration using the faster cell options through the flow had shorter runtimes in the reported comparison without an evident PPA penalty.

These findings reinforce the value of measuring the whole flow. Sign-off optimization may recover timing or power after earlier stages, so exposing more cell choices earlier does not automatically improve the final implementation. Small run-to-run variations can also complicate comparisons.

Build a disciplined evaluation process

For each proposed physical-design change, compare results against a consistent baseline. Track:

  • Power: dynamic, leakage and total power, using a defined workload and switching-activity method.
  • Performance: operating frequency or the relevant timing result.
  • Area: standard-cell area, alongside any applicable design constraints.
  • Runtime: time for individual implementation stages and the complete flow.

Keeping these measures together helps distinguish a meaningful improvement from a change that merely shifts cost between power components, timing and engineering turnaround. A feature that offers a small PPA gain may still be valuable, but its benefit should be considered alongside runtime and the requirements of the target design.

Conclusion: optimize the implemented design, not the setting

Physical implementation presents practical opportunities for ASIC power optimization, including multibit flip-flop merging and selective use of standard cells with different VT and channel characteristics. Yet theoretical advantages do not guarantee better results in a particular flow. Measure power, frequency, area and runtime under controlled conditions, and evaluate dynamic and leakage power separately. This evidence-based approach helps teams choose optimizations that suit their design constraints rather than relying on tool options in isolation.

Access Official Technical PDF

About The Author