Many-Thread Aware Compression in GPGPUs

Compression is a promising technique to increase effective capacity of caches. Due to latency overhead of decompression, most of previous studies applied compression to lower level caches. General-Purpose Graphics Processing Units (GPGPUs) are throughput oriented computing platforms which execute hundreds to thousands of threads, simultaneously. The massive number of threads makes GPGPUs less sensitive to the delay of decompression than CPUs. This makes it possible to exploit compressed L1 data caches, boost performance of GPGPUs. To offset the power overhead of a compressed cache, we use opportunities provided by cache compression to reduce dynamic, static power of the compressed cache. Our evaluation using a set of GPGPU benchmark suites shows that compression improves performance up to 20% with marginal impact on power.