Performance Evaluation of PEACH3: Field-Programmable Gate Array Switch for Tightly Coupled Accelerators

An FPGA switching hub for tightly coupled accelerators (TCA) architecture called PEACH3 (PCI-Express Adaptive Communication Hub ver. 3) is evaluated and its communication speed is analyzed. PEACH3 connects a number of GPUs directly through PCI express Gen3x8 ports. The latency of inter-node GPU-GPU communication of PEACH3 was about 2.8 µ sec which is one third of that of CUDA API with MPI/Infiniband. The bandwidth was about 1.21 times of that of the previous version PEACH2, and 1.54 times of that with MPI/Infiniband for 512KB data transfer. Two application programs: BFS (breadth first search) and CG (conjugate gradient) were implemented with TCA IP and CUDA IP with MPI/Infiniband. The performance of BFS with PEACH3 was 1.16 times better than that with PEACH2, and 1.3 times better than that with MPI/Infiniband for a graph with scale = 15. In CG, for the small matrix (CLASS=S), the PEACH3 achieved 12% better performance than that with PEACH2 and 25% with MPI/Infiniband. However, since the bandwidth of PEACH3 with PCI gen3x8 is smaller than Infiniband with PCI gen3x16, the performance benefit was disappeared for CLASS=A matrix. Through the evaluation, it appears that if the data size is small, using TCA API with PEACH3 is advantageous even for intra-node communication.