A team of U.K. researchers from the University of Warwick’s Performance Computing and Visualization Department and Oxford University’s eResearch Centre have put together a study that takes a look at what may become an increasingly frequent decision high-performance computing sites will have to make: should we install a GPGPU-based system or a more traditional IMB Blue Gene-like supercomputer?
The team’s research will present their work in its entirety at the 1st International Workshop on Performance Modeling, Benchmarking and Simulation of High Performance Computing Systems (PMBS 10) at the upcoming SC10 conference in New Orleans. The paper, entitled“Performance Analysis of a Hybrid MPI/CUDA Implementation of the NAS-LU Benchmark,” compares the performance of a parallel application running across multiple GPU nodes against an InfiniBand-based cluster of AMD processors and an IBM Blue Gene/P.
Their benchmarks show that while the raw performance of GPUs is competitive, often exceeding many general-purpose processor offerings in a device-to-device comparison, “the power-efficiency of the GPU solution is lower than that of the BlueGene/P solution, highlighting the lower levels of sustained performance currently realizable from a GPU solution.”
The study is perfectly timed given last week’s news that the fastest supercomputer in the world is now China’s GPU-based Tianhe-1a, and it is likely that this is the beginning of what may turn out to be a very long debate about just where the future of supercomputer processing hardware is going.