Performance Analysis of Different Accelerators

In recent years, there has been a lot of focus on developing graph analytics algorithms that utilize the high parallelism of GPUs to speed up graph analytics tasks. Meanwhile, the two main GPU manufacturers, NVIDIA and AMD, have made different design decisions that can impact the performance of graph computation. In this chapter, we seek to understand the effects of those decisions through experimentation. Since there is currently no graph analytics software for the ROCm-like platform used by AMD-like GPUs, we have created adGRAPH by porting nvGRAPH, a mature graph analytics library optimized for CUDA, to ROCm-like platforms. AdGRAPH allows us to use AMD-like GPUs to accelerate graph analytics and compare the performance of the two types of GPUs. We tested the performance of several commonly used graph algorithms of varying complexities on two NVIDIA GPUs, A100 and V100, and two AMD-like GPUs developed in China, Z100 and Z100L. Through thorough experiments, we discovered that while NVIDIA GPUs perform better on complex graph analytics algorithms, thanks to their SIMT paradigm, the larger warp size and independent shared memory of AMD-like GPUs make them more efficient than NVIDIA GPUs for graph algorithm implementations with lower branching complexity.