Airi Mini is an entry-level — though powerful — system engineered for deep learning applications.
It comprises two Nvidia DGX-1 deep learning systems (the original Airi had four) with a Pure Storage FlashBlade configured with seven 17TB blades, connected via 100Gbps Ethernet and delivering two petaFLOPS of deep learning performance.
Both Airi and Airi Mini are now available with Cisco's Nexus9000 100GbE switches. Arista 100GbE switches were originally used.
A total of 100GbE is required for AI applications because slower interconnects create a bottleneck between high-performance compute and storage node, he said.
The Airi software includes the Nvidia GPU Cloud deep learning software stack, and the Pure Storage AIRI Scaling Toolkit that allows deep learning models to be trained across multiple DGX-1 nodes. Training performance scales linearly with the number of DGX-1 nodes, Schwartz told iTWire.
And there is room for further improvement. Schwartz noted that the current Airi models use a 15-blade FlashBlade chassis, but Pure is already shipping a 75-blade chassis. In addition, Nvidia recently announced the more powerful DGX-2. "The future's bright," he observed.
"The vast majority of business leaders understand the incredible opportunity presented by AI. Yet implementation complexities often hold their AI initiatives back from moving forward," said Pure Storage's general manager of FlashBlade, Matt Burr.
"While Airi simplifies AI infrastructure for initiatives at any scale, Airi Mini offers a powerful and affordable entry point for organisations to explore and scale as they grow into AI. Airi Mini removes the final barriers for any organisation to leverage AI, and scales as AI workloads grow."
Schwartz explained that "AI is a new workload" for IT organisations – they haven't yet accumulated wisdom in this area, so it is useful for them to be able to buy an "encapsulated" system such as the Airi or Airi Mini.
Nvidia vice-president and general manager Jim McHugh said: "Organisations require infrastructure purpose-built for AI, a system that’s easy to get started with and scales with their efforts.
"Both FlashBlade and DGX-1 are designed to scale seamlessly into the future, enabling customers to grow quickly and build upon their AI success."
On the question of whether AI is better implemented on premises or in the cloud, Schwartz's recommendation is to "do your AI where most of your data is generated".
"Very large datasets have gravity," he observed.
While the actual price wasn't mentioned, Pure said Airi Mini comes "a price point that's accessible for virtually any organisation in any industry".
Disclosure: The writer attended Pure Accelerate as a guest of the company,.