Research on Acceleration Technologies and Recent Advances of Data Center GPUs

Deep learning applications in data centers are significantly computation-intensive and storage-intensive, and require GPUs with high throughput and high energy efficiency. Major manufacturers have successively released a number of high-performance data center GPU accelerators, which comprehensively use a variety of software and hardware optimization technologies to obtain excellent performance. Understanding and using these optimization technologies is of great significance to deeply understand GPUs and fully release their powerful computing capabilities. This paper analyzes and reviews the acceleration technologies and their recent advances used in data center GPU accelerators to improve throughput and energy efficiency, including convolution acceleration, data quantization, tensor/matrix cores, sparsity, high bandwidth and density memory, software stack, and multi-GPU system interconnect. This article also compares and analyzes the computation and storage performance of the latest data center GPU accelerators released by representative vendors such as Nvidia, AMD, Intel and so on.