Selene (supercomputer)

Selene (supercomputer). Selene is a powerful supercomputer system built by NVIDIA and based on its NVIDIA DGX SuperPOD architecture. Named for the Greek goddess of the Moon, it was deployed in mid-2020 at the company's headquarters in Santa Clara, California. The system rapidly gained prominence for its exceptional performance in artificial intelligence workloads and high-performance computing, achieving a top-ten ranking on the prestigious TOP500 list.

Overview

Selene represents a landmark NVIDIA system designed to showcase the capabilities of its integrated NVIDIA DGX SuperPOD platform for AI and scientific computing. As an in-house research instrument, it supports a wide range of projects across deep learning, computational fluid dynamics, and quantum chemistry. The supercomputer's deployment underscored NVIDIA's strategic shift from primarily a hardware supplier to a provider of full-stack computing solutions. Its performance has been validated by major industry benchmarks, including the MLPerf consortium.

Hardware and architecture

The core building block of Selene is the NVIDIA DGX A100 system, each containing eight NVIDIA A100 Tensor Core GPUs and two AMD EPYC central processing units. These nodes are interconnected using a high-speed Mellanox InfiniBand network fabric, specifically HDR InfiniBand, to minimize communication latency. The storage subsystem utilizes a DDN parallel file system for handling massive datasets common in AI training. The entire NVIDIA DGX SuperPOD architecture employs a standardized, modular design, allowing for rapid deployment and scalability.

Performance and rankings

In its debut on the November 2020 TOP500 list, Selene achieved a LINPACK performance of 27.58 PetaFLOPS, securing 7th place globally. It also demonstrated leading efficiency, placing 5th on the Green500 list at that time. The system has consistently posted record-breaking results in AI benchmarks, setting multiple performance records in the MLPerf training competition across categories like natural language processing and recommender systems. These results highlighted the NVIDIA A100's dominance in both high-performance computing and machine learning tasks.

Applications and research

Selene is used extensively by NVIDIA researchers and partners for pioneering work in artificial intelligence. Key projects include developing large-scale language models, advancing autonomous vehicle simulation through platforms like NVIDIA DRIVE Sim, and accelerating drug discovery in collaboration with institutions like the National Institutes of Health. It also runs complex computational physics simulations for climate modeling and materials science. The insights gained from operating Selene directly inform the development of NVIDIA's software stacks, including CUDA and NVIDIA Merlin.

History and deployment

The system was constructed and brought online during the COVID-19 pandemic in mid-2020, a testament to the pre-integrated nature of the NVIDIA DGX SuperPOD design. Its initial deployment at the NVIDIA headquarters in Santa Clara, California provided critical computing resources during a period of high demand for research infrastructure. Selene's performance and design influenced subsequent supercomputer installations worldwide, including the Cambridge-1 system in the United Kingdom. It remains a vital asset for NVIDIA's internal research and development efforts.

Category:Supercomputers Category:NVIDIA Category:Computers introduced in 2020 Category:High-performance computing