An HPC-AI Supercomputer System 1 Set

2026-11-03

Classification of the products to be procured : 1 Nature and quantity of the products to be purchased : An HPC-AI Supercomputer System Set Type of the procurement : Purchase Basi requirements of the procurement : An HPC-AI Supercomputer System consists of parallel computing servers, a parallel file system, login servers, and other hardware and software for stable and efficient operation of the whole system. A Parallel computing servers must satisfy the following conditions. a Parallel computing servers consist of identical compute nodes. b A compute node consists of CPU, GPU, memory, NVMe SSD and other peripherals. Theoretical peak performance of GPU must be equal to or greater than 1 PFLOP for four-bit floating point arithmeti and 15 TFLOP for double-precision floating point arithmetic. The performance of emulation for double-precision floating-point arithmeti is acceptable. d CPU and GPU can cache-coherently access at least 3 TByte of memory operating at a rate of at least 57 TByte/sec. e NVMe SSDs must provide 3. TByte or larger per GPU, with a theoretical peak bandwidth of GB/se or higher. f Each compute node must be connected with a full bisection network interconnect, each link speed must be 80 Gbps or higher per GPU. g The total peak performance of four-bit floating point must be EFLOP or higher. B A parallel file system must satisfy the following conditions. a The parallel file system must be connected to the interconnection network of all compute nodes. b The capacity of user available space must be PByte or larger. The maximum number of inodes must be 20 billion or higher. d Each device of the parallel file system must be swappable without stopping the entire system. e The performance monitoring system for the parallel file system must be installed. C Login servers must satisfy the following conditions. a Login servers must be connected to the interconnection network of all compute nodes. b Each login server must have a network interface for external network at the rate of 10 Gbps or higher, and must be connected to an external network router. The external network router must have a 10 Gbit Ethernet port for external connections. d Login servers must mount the parallel file system of Pegasus and Sieiur supercomputers via the InfiniBand network. D The software available on all the compute nodes must satisfy the following conditions. a The system must be operated by Linux operating system. b For CPU, Fortran, C and C++ including automati parallelization feature and OpenMP must be provided. For GPU, CUDA (including CUDA translator) and OpenACC or OpenMP must be provided. C, C++ and Fortran must be provided. d MPI 3. or later must be provided for parallel processing over all compute nodes. e Highly parallelized numerical and artificial intelligence library must be provided. These libraries must be available in Python. f Batch-job scheduling system to manage all the compute nodes must be provided. E To support the continuous operation of parallel processing servers and the parallel file system which are fully powered, power supply unit capacity, cooling system, floor plan, and all the environments must be carefully considered. Time limit for the submission of the requested material : 1 : 0 2 April, 202 Contact point for the notice : TSURUTA Hitomi Ciref Center for Computational Sciences, University of Tsukuba, 1-1- Tennodai Tsukuba-shi 305-857 Japan, TEL 029-853-6487

Japan

2026-03-26