A Challenge for an Efficient AMI-based Cache System on FPGA Soft Processors

摘要

FPGA soft processors are widely used for embedded applications because of their versatility. High performance cache with fewer resources for FPGA soft processors has become an important issue. Conventional approaches to improve cache performance are following: larger cache capacity, larger associativity and proper line size. These approaches work effectively in many cases, however, cache conflicts occur frequently in some applications because the best cache configuration depends on application characteristics. We propose a cache system on FPGA soft processors which selects the suitable number of cache lines depending on an application to be run. Our proposed system uses Arbitrary Modulus Indexing (AMI) to implement non-power-of-2 cache lines. We evaluate how AMI affects the performance of our proposed system. In evaluation, we confirmed that our proposed cache system can achieve 8.0 speedup on average against conventional cache systems.