Pure Storage has upgraded its combined AI infrastructure offering AIRI with FlashBlade//S storage, making it able to analyze more data faster for better machine learning models and AI insights.
When it was launched in 2018, AIRI (AI-Ready Infrastructure) was a half rack-sized system containing a Pure Storage FlashBlade all-flash array with four Nvidia DGX-1 GPU servers and a pair of 100GbitE switches from Arista. It went through various iterations, with DGX-2 support in 2019, DGX-A100 in 2020, and now FlashBlade//S with the AIRI//S system. En route Mellanox switches were adopted as part of Nvidia’s Spectrum networking.
Pure’s Amy Fowler, strategy and solutions VP for FlashBlade, said: “Traditional approaches to AI infrastructure often result in silos of servers and storage that are either over-spent on capacity or starve AI workloads. With a focus on simplicity and scalability, AIRI//S enables global enterprises to achieve better time to insights and make the most out of their data with AI.”
Charlie Boyle, Nvidia’s VP for DGX systems, said “AIRI//S built on Nvidia DGX systems provides customers with modular, high-performance enterprise infrastructure that scales easily” as demand for integrated, full-stack computing optimized for AI continues to grow.
Pure says AIRI//S systems can be set up, deployed, and managed quickly as an end-to-end AI pipeline solution – a dedicated AI box set. It can scale storage and AI compute capabilities. AIRI//S, compared to existing AIRI systems, is a more powerful, high-performance AI infrastructure in the same footprint. It delivers the FlashBlade//S advantages of higher density, reduced power consumption, and better datacenter efficiency. AIRI//S can also be operated and consumed on a subscription basis through the Evergreen//One and Evergreen//Forever services.
AIRI//S, meaning FlashBlade//S, does not yet support Nvidia’s MagnumIO GPUDirect Storage (GDS). Pure CTO Rob Lee said that the total work time for an AI job includes searching a source dataset for the desired subset, extracting it and then sending it to the GPUs. The search and extract process can take far longer than the data transfer to the GPUs. Lee said FlashBlade//S shortens that time so that the overall job time is lowered by more than just having a faster data transfer time, which is what GDS provides.
AIRI is a modular system with the ability to swap in faster components such as network switches, GPUs, and storage hardware and software. FlashBlade//S and its PurityFB OS will support GDS next year. AIRI//S will adopt it as a result, and then AIRI//S will be an even faster system.