Parallel Architecture
Albert Chun Chen Liu, Oscar Ming Kin Law · 2021
This chapter describes several popular parallel architectures, Intel central processing unit (CPU), NVIDIA Graphics Processing Unit (NVIDIA GPU), Google Tensor Processing Unit, and Microsoft NPU. Intel CPU is designed for general computation and NVIDIA GPU is targeted for graphical display. Traditionally, the CPU is targeted for general-purpose computation. It is evolved from Single Instruction Single Data architecture to Single Instruction Multiple Data one to support parallel processing. Intel Scalable Xeon processor upgrades to Intel Ultra Path Interconnect mesh architecture to resolve the latency and bandwidth issues. Caching and Home Agent are integrated together to form a new Combined Home Agent to resolve the memory bottleneck. Google also develops new software architecture for cloud computation. The Catapult fabric is derived from the single thread vector processor architecture with Matrix-Vector Multiplier, the multifunction unit, and vector arbitration networks.