PEZY-SC3: A MIMD Many-core Processor for Energy-efficient Computing

Abstract: PEZY-SC3 is a highly energy- and area-efficient processor for supercomputers developed using TSMC 7nm process technology. It is the third generation of the PEZY-SCx series developed by PEZY Computing, K.K. Supercomputers equipped with the PEZY-SCx series have been deployed at several research centers and are used for large scale scientific calculations. PEZY-SC3 outperforms previous PEZY-SCx and other processors in terms of energy and area efficiency. To achieve high efficiency, PEZY-SC3 employs a MIMD many-core, fine-grained multithreading, and non-coherent cache, focusing on applications involving high thread-level parallelism. Our MIMD many-core-based architecture achieves high efficiency while providing higher programmability than existing architectures based on specialized tensor units with limited functionality or wide-SIMD. Another key point of this architecture is to achieve both high efficiency and high throughput without using complex and expensive units such as out-of-order schedulers. Moreover, our novel non-coherent and hierarchical cache system enables high scalability on many-core without compromising programmability. The energy efficiency of a system equipped with PEZY-SC3 is approximately 24.6 GFlops/W, and it ranked 12th in the Green500 (November 2021), which measures the energy efficiency of supercomputers. In terms of processor architecture, all the systems ranked higher than the PEZY-SC3 system are equipped with NVIDIA A100 or Preferred Networks MN-Core, and thus PEZY-SC3 is the third-ranked processor after them. While A100 and MN-Core achieve high energy efficiency with tensor units specialized for specific functions, PEZY-SC3 does not have such specialized tensor units and thus has higher programmability.