With the shift from scaling frequency to scaling the number of cores, efficiency becomes a primary design consideration. The ability to scale the number of cores while pushing back the memory and power walls with small increases in die size will require significant improvements in cache efficiencies. This paper provides strategies to improve L2/L3 data cache efficiencies by coupling voltage scaling with flexible cache management policies. Specifically, we propose a framework that encompasses i) off-line creation of a voltage-area profile, ii) on-line measurement of cache line utilization to drive voltage scaling, and, iii) changing the placement function to match the voltage-scaled area and the program-phase cache footprint. The proposed techniques were applied to several benchmarks resulting in performance efficiencies doubling, energy efficiencies improving by 10% (relatively) with a 10% improvement in Energy Delay Product.
展开▼