Highlights
- Plans to build with 4th Gen Intel Xeon Scalable processors (formerly codenamed Sapphire Rapids), NVIDIA H100 Tensor Core GPU with PCIe and 51TFlops of extreme performance, and Intel Optane persistent memory (codenamed Crow Pass) will strongly drive big data and AI
- The world’s first system with NVIDIA H100 GPUs connected via PCIe Gen5
- The first system announced in Japan that will utilize NVIDIA Quantum-2 400Gb/s InfiniBand networking
Tsukuba, Japan – December 22, 2022 – The Center for Computational Sciences at the University of Tsukuba announces plans for a new big memory supercomputer Pegasus manufactured by NEC. The new supercomputer will be one of the first systems in the world to introduce 4th Gen Intel Xeon Scalable processors (formerly codenamed Sapphire Rapids), Intel Optane persistent memory (codenamed Crow Pass), and the NVIDIA H100 Tensor Core GPU with 51TFlops of breakthrough acceleration.
NVIDIA H100 PCIe GPU is the latest GPU announced in March, 2022 with a theoretical peak performance of 51 TFlops in double precision floating point operations. Pegasus is the first to offer the latest high-bandwidth PCIe Gen5 connection to H100 GPU. Networking utilizes the latest NVIDIA Quantum-2 400 Gb/s InfiniBand networking platform with advanced in-network computing acceleration engines. Utilizing the world’s newest technology such as the Memory bandwidth and I/O bandwidth increased by DDR5 and PCIe Gen5, the University of Tsukuba will strongly support HPC, big data and AI.
To drive big data analytics and ultra-large scale AI, not only high computing performance is required, but also high-bandwidth large memory and ultra-fast storage systems; Pegasus has introduced H100 GPUs, DDR5 memory, and persistent memory to solve this problem. The H100 GPU is 2.7 times faster in computing performance than the current A100 GPU, and DDR5 memory provides twice the memory bandwidth of conventional DDR4 memory. The Center of Computational Sciences has been offering solutions with Intel Xeon processors and Intel Optane persistent memory since 2020 to research partners. Based on previous generations of Intel Optane persistent memory, allowing direct access to persistent data structures has potential to significantly improve application performance. In addition, unlike DRAM, persistent memory does not require refresh to retain data, thus dramatically reducing power consumption compared to DRAM of the same capacity. Moreover, memory expansion with persistent memory enables calculations with memory sizes that exceed DRAM capacity, and the performance of computational science applications remains largely unchanged.
The effectiveness of compute accelerators (GPUs) in HPC, data-driven science, and AI-driven science has already been demonstrated, and Pegasus has been designed to incorporate state-of-the-art CPUs, GPUs, memory, and persistent memory, connected via a state-of-the-art network.
Pegasus consists of 120 compute nodes, and the overall theoretical peak performance of the GPU part alone exceeds 6.1 PFlops. The Center for Computational Sciences at the University of Tsukuba, as a Joint Usage / Research Center, will make Pegasus available to users around the world through various usage programs such as Interdisciplinary Collaborative Use, HPCI Collaborative Use, and General Use, thereby contributing to the further development of computational science.
“Big data and AI applications are one of the most important research topics in addition to the high performance computational sciences which we have been focusing on so far. Introducing a new machine with a large-capacity memory system and high-performance AI processing should be our new tools to expand our research field toward excellent data science,” said Taisuke Boku, Director of the Center for Computational Sciences, the University of Tsukuba.
“Data-driven and AI-driven science requires large memory size and storage performance, but memory capacity per CPU core has decreased for a decade,” said Osamu Tatebe, professor at the Center for Computational Sciences, the University of Tsukuba. “That is why we introduce NVIDIA H100 PCIe GPU, DDR5, and persistent memory in compute nodes to accelerate large-scale data analysis and big data for better cost performance, power consumption, and application performance. Also, we foster new fields of large-scale data analysis, new applications of big data AI, and system software research.”
“NEC is very honored to announce that the Center for Computational Sciences at the University of Tsukuba adopted NEC LX-Series as their new supercomputer,” said Yasuo Mogaki, Managing Director, NEC Advanced Platform Division. “This system is World’s first system that delivers revolutionary acceleration applying Intel’s next generation Xeon CPU, Optane memory and NVIDIA’s H100 GPU. NEC will contribute to advanced computational science by combining our technological capabilities.”
“We are very excited that the University of Tsukuba plans to build their new supercomputer with the upcoming 4th Gen Intel Xeon Scalable processor and Optane persistent memory 300 series, showcasing the value that Intel Architecture brings to high performance computing” said Don Cunningham, Vice President & General Manager, Data Center and AI Group at Intel.
“As supercomputing enters the era of exascale HPC and AI, NVIDIA enables researchers to tackle massive challenges using NVIDIA H100 GPUs,” said Ian Buck, vice president of Hyperscale and HPC at NVIDIA. “The selection by the University of Tsukuba will enable researchers to deliver the same AI performance with 3.5x more energy efficiency and 3x lower total cost of ownership.”
“It is our great pleasure to support the innovative Pegasus project at Tsukuba University. We look forward to working with the team at Tsukuba University towards better integrating emerging memory architectures and node-local persistent storage solutions with global shared storage” said Robert Triendl, President of DataDirect Networks Japan, Inc.
Planned Specifications
|
System name |
Pegasus |
|
Manufacture |
NEC |
|
Total performance |
> 6.1 PFlops |
|
Number of nodes |
120 |
|
Interconnects |
Full bisection fat-tree network interconnected by the NVIDIA Quantum-2 InfiniBand platform |
|
Parallel file system |
7.1PB DDN EXAScaler (40 GB/s throughput) |
Compute node
|
CPU |
4th Gen Intel Xeon Scalable processor formerly codenamed Sapphire Rapids (48c) |
|
GPU |
NVIDIA H100 Tensor Core GPU with PCIe Gen5 (51 TFlops in FP64 Tensor Core, 80GB HBM2E, 2 TB/s) |
|
Memory |
128GiB DDR5 (282 GB/s) |
|
Persistent memory |
Intel Optane persistent memory (codenamed Crow Pass) |
|
SSD |
2 x 3.2TB NVMe SSD (7 GB/s) |
|
Networking |
NVIDIA Quantum-2 InfiniBand platform (200 Gb/s) |
About the Center for Computational Sciences, the University of Tsukuba
For more information, go to www.ccs.tsukuba.ac.jp