high performance computing on graphics processing units: hgpu.org

In recent studies, phase changing memory (PCM) has shown promising energy efficiency for systems with a modest level of parallelism. But it remains an open question whether it can benefit GPU-like massively parallel systems. This work conducts the first systematic investigation into this question. It empirically shows that contrary to the promising results shown before on CPU, the previous designs of PCM-based memory result in significant degradation to the energy efficiency of GPU computing. It reveals that the fundamental reason comes from a mismatch between those designs and the massive parallelism in GPU. It further shows that fixing the mismatch requires innovations in both hardware and software support. It introduces a set of new hardware features and a novel compiler-directed data placement scheme to address the issues. Working hand-in-hand, they tap into the full potential of hybrid memory for GPU, yielding 15.6% and 40.1% energy saving on average comparing to pure DRAM and pure PCM respectively, and keeping performance loss less than 3.9%.