Logic Filter Cache for Wide-VDD-Range Processors
Paper i proceeding, 2016
Wide-VDD-range processors offer high energy efficiency for varying embedded workloads. But reducing the VDD of the cache as aggressively as the VDD of the CPU logic is not straightforward, since standard 6T SRAMs cease to operate at lower VDDs. We implement a data and instruction filter cache, using logic cells located in the CPU VDD domain, to permit the level-1 (L1) cache to be reliably powered at a higher SRAM VDD. On top of eliminating many energy-wasting L1 cache accesses, the filter cache reduces the total number of executed cycles. Furthermore, the filter cache can be reconfigured as CPU VDD is reduced, to filter out an increasing proportion of cache accesses. We evaluate our approach using a 65-nm 1.2-V low-leakage CMOS process, with a minimal CPU and SRAM VDD of 0.4 and 0.95V, respectively. Assuming 16kB+16kB L1 caches and 256B+256B filter caches, introducing the filter cache reduces the total cache access energy by 71% at 1.2V and 87% at 0.4V at an area overhead which is 13% of the L1 cache area.