This article details Japan's sustained investment in custom floating-point accelerators developed by Pezy Computing KK, highlighting their role as an energy-efficient alternative to GPUs for High-Performance Computing (HPC) and AI. Driven by funding from NEDO, Pezy has evolved its "SC" series of chips over a decade, culminating in the latest Pezy-SC4s. This accelerator, fabricated on a 5nm process, features 2,048 custom RISC Processor Elements (PEs), HBM3 memory, a sophisticated cache hierarchy, and integrated RISC-V cores for self-contained operation. The architecture leverages a Single Program, Multiple Data (SPMD) approach with fine-grained multithreading for high computational efficiency and flexibility. Performance comparisons, particularly in genomics workloads, suggest Pezy accelerators can rival or exceed NVIDIA GPUs in FP64 performance per watt. Japan's continued commitment to Pezy Computing is viewed as a strategic hedge against potential GPU supply chain issues and a means to retain critical domestic hardware development expertise.
Japan's Strategic Investment in Pezy Custom Floating Point Accelerators
0:00 Introduction to Pezy Computing: Japan is actively investing in custom floating-point accelerators developed by Pezy Computing KK, a company focused on highly energy-efficient math accelerators for HPC and AI. This initiative, funded by Japan's New Energy and Industrial Technology Development Organization (NEDO), aims to provide an alternative to traditional GPUs.
1:05 Evolution of Pezy-SC Accelerators: Pezy Computing has developed a line of "SC" (Super Computer) accelerators since 2012.
Pezy-SC3s (2021): A cut-down version of SC3, possibly for exploring cost-effective node designs.
5:10 Pezy-SC4s Unveiled (2025 Release): The latest iteration, fabricated on TSMC's 5nm process, features:
Core Count: 2,048 PEs, capable of 16,384 threads.
Clock Speed: 1.5 GHz.
Memory: 96 GB HBM3 with 3.2 TB/sec bandwidth.
Cache: 64 MB L3 cache.
Power: Estimated 600 watts.
Precision Support: Adds BF16 to existing FP64, FP32, FP16.
Host Cores: Integrates four "Rocket" open-source RISC-V CPUs running Linux, eliminating the need for an external x86 host for OS management.
8:15 Architectural Philosophy (SPMD): Pezy employs a Single Program, Multiple Data (SPMD) architecture, a variant of MIMD, which allows for explicit thread and data scheduling. This approach simplifies cores by removing complex features like branch predictors and out-of-order execution, enabling a higher core count. Fine-grained multithreading helps hide memory latency.
10:05 SC4s Cache Hierarchy: PEs have 4KB L1 instruction/data cache and 24KB scratchpad. PEs are grouped into "villages" (4 PEs), then "cities" (4 villages sharing 32KB L2 instruction/64KB L2 data), then "prefectures" (18 cities, 16 active), and finally "states" (8 prefectures sharing 64MB L3 cache).
12:10 System Integration: A proposed system board features an AMD Epyc 9555P host, a 400 Gb/sec NDR InfiniBand controller, and four SC4s accelerators connected via PCIe. Plans exist for a 90-node test system with 8.6 petaflops FP64 performance.
13:20 Software Stack and AI: Pezy Computing has developed its own software stack, including an AI stack supporting PyTorch. Major AI models like Google Gemma3, Meta Llama3, Alibaba Qwen2, Stable Diffusion 2, and Hugging Face HuBert have been ported to the SC4s.
14:30 Performance Benchmarking and Efficiency:
GATK Genomics: A quad of Pezy-SC3 chips processed genomics analysis 2.25X faster per sample than eight NVIDIA H100 GPUs (33 minutes vs. 37 minutes). The SC4s is projected to be ~2.8X better than an H100 on GATK.
Flops per Watt (FP64): Pezy-SC3 delivered 41.9 gigaflops/watt. The SC4s is estimated at 41 gigaflops/watt, with the future SC5s targeting 45.8 gigaflops/watt. This is competitive with NVIDIA H200 (47.9 gigaflops/watt) and superior to B200 (33.3 gigaflops/watt) at FP64.
16:30 Strategic Importance for Japan: Japan's investment in Pezy Computing is a strategic move to:
Hedge against supply chain risks: Ensure access to high-performance accelerators even if GPUs are scarce or restricted.
Retain domestic expertise: Maintain a national skillset in designing advanced math accelerators.
17:35 Future Outlook: While the upcoming FugakuNext supercomputer will heavily feature NVIDIA GPUs, there is a strong rationale for including Pezy accelerators in the future, with a Pezy-SC5s and subsequent generations expected.