Closing the Gap between ASIC and Custom Power

David Chinnery
(Professor Kurt Keutzer)

We have carefully examined the speed gap between ASIC and custom processors. Typical ASIC designs are a factor of x3 to x8 slower than comparable custom designs. The contributing factors to this gap are microarchitecture, timing overhead, logic style, logic design, cell design and wire sizing, layout, and process variation and accessibility. With careful attention to reduce these contributions, some ASICs can close the speed gap to only a factor of x2 to x3 [1].

Power is also a significant concern to designers. The heat dissipated by a chip greatly impacts the cost of packaging and cooling methods (ranging from passive dissipation to refrigeration). The power consumption limits battery life and adds to the cost of running an instrument. Particularly for low cost embedded applications, power consumption directly limits the maximum performance.

Custom logic styles such as dynamic domino logic are perceived as being higher power consumption than static logic. However, as dynamic circuits achieve higher maximum speeds, they can be downsized more and can reduce the supply voltage to meet slower speed targets. Custom logic styles typically have less capacitance in the PMOS pull-up network of gates, due to alternative logic styles and skewed drive strengths. As the gate loads are smaller, the drive strengths required are smaller, doubly impacting the dynamic power consumption of switching capacitances within the circuit.

Given equivalent speed targets for ASIC and custom methodologies, custom approaches can exploit their speed advantage to substantially reduce power consumption. We quantify the contribution of each factor to the power gap between ASIC and custom designs. We then illustrate approaches to reducing ASIC power consumption.

[1]
D. G. Chinnery and K. Keutzer, Closing the Gap between ASIC and Custom: Tools and Techniques for High-Performance ASIC Design, Kluwer Academic Publishers, ISBN 1-4020-7113-2, May 2002.

Send mail to the author : (chinnery@eecs.berkeley.edu)


Edit this abstract