high performance computing on graphics processing units: hgpu.org

Preserving memory locality is a major issue in highly-multithreaded architectures such as GPUs. These architectures hide latency by maintaining a large number of threads in flight. As each thread needs to maintain a private working set, all threads collectively put tremendous pressure on on-chip memory arrays, at significant cost in area and power. We show that thread-private data in GPU-like implicit SIMD architectures can be compressed by a factor up to 16 by taking advantage of correlations between values held by different threads. We propose the Affine Vector Cache, a compressed cache design that complements the first level cache. Evaluation by simulation on the SDK and Rodinia benchmarks shows that a 32KB L1 cache assisted by a 16KB AVC presents a 59% larger usable capacity on average compared to a single 48KB L1 cache. It results in a global performance increase of 5.7% along with an energy reduction of 11% for a negligible hardware cost.