Tera-scale Computing: Ultra-low Voltage Technologies for Energy-efficient Special-purpose Hardware Accelerators

Tera-scale Computing: Ultra-low Voltage Technologies for Energy-efficient Special-purpose Hardware Accelerators

This article describes ultra-low voltage design techniques and learnings from a video motion estimation engine fabricated in 65nm CMOS technology. This chip is targeted for special-purpose, on-die acceleration of sum of absolute difference (SAD) computation in real-time video encoding workloads on power-constrained mobile microprocessors. Various datapath circuit innovations within the accelerator improve energy efficiency for SAD calculations at nominal supplies, while ultra-low voltage optimizations enable robust circuit operation for further efficiency gains at ultra-low supply voltages, with minimal impact on nominal supply performance. Silicon measurements of the accelerator demonstrate performance of 2 GHz at the nominal supply voltage of 1.2 V, with scalable performance of up to 2.4 GHz at 1.4 V, 50°C. Robust, ultra-low voltage, optimized circuits enable operation measured down to 230 mV (sub-threshold). Across this wide range of operational supplies, maximum energy efficiency of 411 GOPS/W or 12.8 macro-block SADs/nJ is achieved by operating the accelerator at a near-threshold voltage of 320 mV, for 23 MHz frequency and 56 µW power consumption. This represents a 9.6X higher efficiency than at the nominal 1.2 V operation. 

Read this article and keep your skills sharp! Actively continuining your learning is an excellent way to prepare for your next innovation even while you work on today's project.


Update your feed preferences