Tokyo Tech Deploys TSUBAME4.0 Supercomputer, Boosting Computational Power

The Global Scientific Information and Computing Center (GSIC) at Tokyo Tech started operation of its next-generation supercomputer, TSUBAME 4.0, on April 1.

From “Everyone’s Supercomputer” to “More of Everyone’s Supercomputer”

Since TSUBAME1.0 started operation in April 2006, the TSUBAME series of supercomputers at Tokyo Tech has long supported research and development in industry, academia, and government, both in Japan and overseas, as “everyone’s supercomputer.” The latest system, TSUBAME4.0, is now in operation at Suzukakedai Campus and makes extensive use of its GPUs (graphics processing units)[1], a feature of the TSUBAME series, and has achieved a 5.5 to 20-fold improvement in performance and ease of use while maintaining the continuity with the TSUBAME3.0 system.
In addition, for the first time in the TSUBAME series, the public was invited to submit rack designs. The design was inspired by the various possibilities and applications of the new supercomputer, and depicts four swallows freely drawing flowing lines as they spread their wings over an endless world.

The use of the TSUBAME series is open not only to Tokyo Tech’s researchers, but also to extramural and corporate researchers through joint use programs such as the Joint Usage and Research Center for Interdisciplinary Large-scale Information Infrastructure (JHPCN) and the Innovative High Performance Computing Infrastructure (HPCI). The aim is to continue to make it “more of everyone’s supercomputer” and to have it be used on a daily basis.

TSUBAME4.0 configuration and performance

TSUBAME4.0 boasts high performance in 64-bit double precision[2], which is mainly used in scientific and technical calculations, as well as 952 petaflops[3] in 16-bit half precision[4], which is used in AI. It is equivalent to the second largest supercomputer among existing supercomputers at research and educational institutions in Japan, following “Fugaku”.

Table 1. TSUBAME4.0 configuration

Number of computation nodes/units

HPE Cray XD665, 240 units

Below, per unit

Processor

AMD EPYC 9654 (96 cores)×2 units

Main memory

DDR5-4800 768 GiB

GPU

NVIDIA H100 SXM5 94GB HBM2e×4 units

High-speed network

InfiniBand NDR200 200 Gb×4 units

Local storage

1.92 TB NVMe SSD

Shared storage

HPE Cray ClusterStor E1000

Table 2. TSUBAME4.0 system performance

Total computing performance

Double-precision

66.8 petaflops (5.5 times that of TSUBAME 3.0)

Half-precision

952 petaflops (20 times higher than TSUBAME 3.0)

Shared Storage Capacity

Hard disk section 44.2 PB (2.8 times larger than TSUBAME 3.0)
SSD section 327 TB (new)

TSUBAME4.0 unveiling ceremony and future prospects

To commemorate the start of operations, GSIC held a TSUBAME4.0 unveiling ceremony and tour on the Suzukakedai Campus on April 18, attended by about 100 guests.

The unveiling ceremony began with opening remarks by GSIC Director Toshiya Ito, followed by congratulatory speeches from the guests of honor. Following an overview of TSUBAME4.0 by Professor Toshio Endo of GSIC, two academic lectures were given by Professor Yutaka Akiyama of the Department of Computer Science, School of Computing Science and Engineering, on “Supercomputers Accelerate the Trend of Medium Molecular Drug Discovery,” and by Professor Rio Yokota of GSIC on “Large-scale Language Model Swallow by Tokyo Tech and AIST (National Institute of Advanced Industrial Science and Technology).” The lecturers introduced the cutting-edge research that has been achieved through the TSUBAME series, and expressed their expectations for a significant acceleration of research using TSUBAME4.0.

As Tokyo Tech and Tokyo Medical and Dental University will merge and become Institute of Science Tokyo (Science Tokyo) in October this year, President Kazuya Masu said, “TSUBAME4.0 will play an important role not only in the existing research of Tokyo Tech, such as meteorology, physical science, materials science, computational chemistry, image processing, language processing, and artificial intelligence processing, but also in the development of academic fields such as AI, medical and dental sciences, and geo-environmental sciences.” TSUBAME4.0 will contribute as an indispensable computational resource infrastructure to the “development of convergence science,” which is the goal of Science Tokyo.