A year ago, Pure Storage introduced AIRI [AI-Ready Infrastructure] in partnership with NVIDIA, which combined the NVIDIA DGX-1 AI supercomputer and the Pure FlashBlade into an enterprise ‘AI-in-a-Box.’ They followed that up in May with the announcement of the AIRI Mini at their Accelerate event. AIRI Mini is a smaller version, with less compute capacity, with its goal being to democratize AI within the enterprise. Now they are addressing the other end of the AI market with the new Hyperscale AIRI. They chose not to call it AIRI Maxi. But they could have.
“This year, we are doing things on the right side of the portfolio, the high end of the product line,” said Brian Schwarz, VP of Product Management for FlashBlade at Pure Storage. “It made sense to start on the left side. When customers adopt AI they start small, with AIRI and AIRI Mini. We had a good year with AIRI and have lots of customers. Now we are building out Hyperscale AIRI. The biggest systems look more like supercomputers and we wanted to use AIRI system to provide a bigger system with supercomputing capabilities.”
Hyperscale AIRI is a partnership between Pure Storage, NVIDIA, and Mellanox, which has since been acquired by NVIDIA, although the deal has not yet closed. Hyperscale AIRI offers NVIDIA DGX-2 systems as well as DGX-1, which the earlier versions could not. The three new Hyperscale AIRI models – two for the DGX-2 and one for the DGX-1 – bring the family to five models in total.
With NVIDIA NGC software container registry and AIRI scaling toolkit, data scientists can begin building applications with containerized AI frameworks. Integration with Kubernetes and Pure Service Orchestrator gives the AI-infrastructure cloud-like elasticity. Since Hyperscale AIRI requires a network that can support high-performance computing environments, Mellanox’s contribution is the industry’s lowest latency and highest bandwidth network. Both Infiniband and Ethernet fabrics are available as interconnect options.
Schwarz said that while Hyperscale AIRI is a hyperconverged infrastructure [HCI] product, it’s a fundamentally different kind of product from those offered by HCI vendors.
“Typical HCI is more about standard business processes – running VMs,” he said. These are for AI – neural network training. You need a different type of architecture for the servers and the storage. If you run machine learning on a traditional Intel server it would take 1000 times longer, as these AI processes would literally take months to complete. These take days instead.”
Hyperscale AIRI has real-world deployments, but none of the present customers are publicly referenceable.
Pure Storage only sells through channel partners, but not many partners are likely to sell Hyperscale AIRI.
“First, you need to be a partner of all three vendors,” Schwarz stated. “Mellanox is still a separate company, and some partners may not have that standalone relationship. A partner also needs to have skill and competency around DGX-2, which is a more complex device. Partners who work with this will be on Chapter Three or Chapter Four of their AI journey. If a partner hasn’t already done simpler AI deployments, it’s hard to imagine how they could do Hyperscale ones.”
The other part of the Pure Storage announcement is likely to have a much more significant channel component. They announced FlashStack for AI, a souped-up version of FlashStack built jointly with Cisco and NVIDIA, which Pure is touting as bringing AI within reach for every enterprise with traditional analytics. It combines Pure’s FlashBlade with the Cisco UCS C480ML server and NVIDIA Tesla V100 eight-way GPUs.
“We have a very broad network of joint partners with Cisco, and there will be a very big channel component for this,” Schwarz said.