Singapore’s National Supercomputing Center has announced that they will be using AMD’s EPYC Milan and NVIDIA’s A100 GPUs to power their upcoming 10 PFLOP supercomputer.
Singapore’s NSCC Supercomputer Powered By 100,000 AMD EPYC Milan CPU Cores & NVIDIA A100 GPUs To Feature 10 PFLOPs Compute Power!
The supercomputer will be utilized by the NSCC (National Supercomputing Center) for research institutes, government agencies, and companies, reports The Register. The Supercomputer has received $200m in investment directly from the Singapore government (March 2019).
The next generation national supercomputer for Singapore will be a green, warm water-cooled system – one of the first known deployments of such a system in a tropical environment. When operational the supercomputer is expected to provide an aggregate of up to 10 PFLOPS of raw compute power and is eight times more powerful than the current ASPIRE1 supercomputer.
ASPIRE1, which was commissioned in 2016, has been running at near full capacity in support of local advanced research that requires high-end computing resources. The new system is the first in a series of supercomputers that will be deployed in phases from now till 2025 to expand and upgrade Singapore’s high-performance computing (HPC) capabilities for the research community here.
The new system, which will be 8X faster compared to NSCC’s existing pool of HPC resources, will expand and augment ongoing research efforts by enabling tools such as artificial intelligence (AI) and deep machine learning to optimize modeling, simulation and even software simulation for quantum computing.
The Supercomputer itself will be based on the latest CPU / GPU architectures from NVIDIA and AMD. It will be based on the HPE Cray EX product lineup and feature a total of 900 computing nodes. The exact system features over 100,000 cores based on AMD’s Zen 3 core architecture. It comprises the AMD EPYC 7003 ‘Milan’ lineup of server processors which feature up to 64 cores & clock speeds hitting almost 3.70 GHz.
NSCC’s supercomputer will also pack 352 NVIDIA A100 Tensor Core GPUs which add another quarter-million GPU cores to the system. We know that the Tensor A100 offers a peak compute power of 19.5 TFLOPs so the system is getting at least 6.8 PFLOPs of compute power out of the GPUs alone whereas the rest of the 3.2 PFLOPs of compute power is driven by the AMD 3rd Gen EPYC Milan chips. The supercomputer will also feature multiple 100 Gbps links so that it could work with SingAREN-Lightwave Internet Exchange from the Singapore Advanced Research and Education Network.
There is no official name for the supercomputer yet but it is expected to be commissioned and become operational by early 2022.