The firm today also unveiled an alternative platform aimed at the more mainstream enterprise market. This is based on its FlashStack collaboration with Cisco, using UCS servers and switches combined with Nvidia GPUs and Pure Storage FlashBlades.
The first iteration of Pure Storage’s AIRI (AI-Ready Infrastructure) launched a year ago, based on Nvidia’s DGX-1 GPU-accelerated Intel boxes and followed later by a more modest AIRI Mini configuration.
Hyperscale AIRI takes the portfolio in the opposite direction and is designed to eliminate the challenges that prevent organisations from deploying AI at scale, Pure Storage said.
It can scale out to multiple racks of Nvidia’s DGX-1 and the more powerful DGX-2 systems, with Infiniband and Ethernet fabrics available as interconnect options, plus Pure Storage FlashBlades providing the storage layer.
The platform offers data scientists an AI infrastructure with cloud-like elasticity thanks to Nvidia’s NGC software container registry, the AIRI scaling toolkit and integration with Kubernetes and Pure Service Orchestrator.
The appearance of Mellanox networking in this platform is notable because Nvidia is acquiring the firm precisely because of its high performance connectivity expertise.
FlashStack for AI
FlashStack has been around for a few years as a converged infrastructure solution using Pure Storage all-flash arrays, Cisco UCS servers and Nexus switches. The two partners have updated this to FlashStack for AI by adding Nvidia GPUs to a Cisco UCS C480ML server and upgrading the storage to FlashBlades.
Cisco said that with these upgrades, FlashStack for AI enables enterprise customers to extend their existing infrastructure to support AI/ML workloads without having to add new infrastructure silos.
Hyperscale AIRI is available worldwide now, while FlashStack for AI is due to ship at the end of April. No details on pricing have been offered, but the original AIRI was estimated to cost somewhere upwards of $1m.
This article was published first on The Register.