Cascaded Cache Based on Recently Used Order for Latency Optimization for IoT
Juhee Choi, Heemin Park · Journal of Computing Science and Engineering · 2021
The load-to-use latency of the L1 cache is one of the main factors to determine the performance of low power processors. As the clock frequency competitions become severe even in Internet of Things (IoT) devices, the access cycle to the L1 cache was increased to meet the timing constraints. Because previous studies assumed that every access to the data cache has the same latency, the latencies tend to be longer for higher performance. We propose a latency optimization method that uses a cascaded cache based on a recently used order that has several small banks with various latencies instead of one whole data cache with constant latency. In our cache, each bank contains the cache block according to its recently used order. The experiments show that the performance of our proposal is improved by 23.0%, with a 1.3% reduction in the dynamic energy consumption on average.