Implementation and Evaluation of NAS Parallel CG Benchmark on GPU Cluster with Proprietary Interconnect TCA

We have been developing a proprietary interconnect technology called Tightly Coupled Accelerators (TCA) architecture to improve communication latency and bandwidth between accelerators (GPUs) over different nodes. This paper presents a Conjugate Gradient (CG) benchmark implementation using the TCA and results of performance evaluation on the HA-PACS/TCA system, which is a proof-of-concept GPU cluster based on the TCA concept. The implementation is based on the CG benchmark in NAS Parallel Benchmarks, and its parallelization is achieved by a two-dimensional decomposition of matrix data. The TCA utilization improves the communication performance compared with the implementation with MPI/InfiniBand utilization for small size benchmark classes. This study also shows that the CG implementation with the two-dimensional decomposition is more suitable for the TCA utilization than a CG implementation with a one-dimensional decomposition to make use of the interconnect.