-
For the benchmarks used, the energy reduction achieved by using a Filter Cache (as opposed to a Nominal 2-level cache) is of the order of 37-45% with a marginal performance penalty of 2-5%. The corresponding temperature values of the I-Cache are reduced by 7-10%. Increasing the size of the Filter Cache (256 bytes from 128 bytes) reduces the energy values further by 2-5% but there is negligible effect on the temperature values.
-
HotSpot Cache reduces the energy further by 18-22% and temperature values by 4-6%. This is much lower than the proclaimed 52% in the original paper which is probably because the benchmarks used are not multimedia applications (comprising of distinct phases of different sets of ‘hot’/frequently accessed basic blocks). Increasing the size of the filter cache (L0) in the HotSpot architecture reduces energy values by another 5 10%. The temperature values tend to fall in the same range for both Filter Cache and HotSpot with the values for the HotSpot Cache being marginally lower.
-
In any cache-architecture, having 4-way or 8-way associativity provides the best results in terms of energy and temperature while the values increase for both direct-mapped and fully-associative configurations.
-
The LRU and FIFO replacement policies tend to have similar energy and temperature values for any cache configuration. The ‘random’ replacement policy tends to consume more energy and cause a higher temperature in the I-Cache.