GPU Reference Architecture

State of the Art PB-Scale Infrastructure for GPU-Intensive and Storage Intensive AI Workloads

As AI becomes increasingly pervasive in our world – across all industries and verticals – organizations need a simpler way of providing modern IT infrastructure. Just as traditional compute (CPU-based) infrastructure is ill-suited for AI workloads, legacy storage architectures are no longer adequate in meeting the I/O demands of GPU equipped systems.

This document details the VAST Data Universal Storage reference architecture for machine learning (ML) and artificial intelligence (AI) workloads and contains benchmarking results obtained in partnership with NVIDIA. This reference design is implemented using VAST Data’s LightSpeed all-flash storage system, four NVIDIA DGX™ A100 systems, and NVIDIA® Mellanox® Quantum™ InfiniBand and Spectrum Ethernet switches.

Download this white paper now and learn:

  • How training and inferencing workloads perform with VAST Data Universal Storage and NVIDIA DGX A100 systems
  • How to achieve high and predictable performance to meet any machine-learning and deep-learning workload requirements
  • How to design and deliver high-throughput storage infrastructure, without the need for complex parallel file systems