A jointly engineered ‘AI-in-a-Box,’ AIRI allows customers to cut simulation training time by 4x, in a much tinier footprint than it took before.
Today, at GTC 2018, NVIDIA’s GPU Technology Conference, Pure Storage and NVIDIA are announcing AIRI [AI-Ready Infrastructure], which integrates Pure FlashBlade and the NVIDIA DGX-1 AI supercomputer into an enterprise ‘AI-in-a-Box.’ AIRI will initially go to market through a select channel of highly-focused joint partners of both companies.
“AIRI is the industry’s first complete AI-ready infrastructure, co-architected by Pure and NVIDIA,” said Matt Burr, GM of FlashBlade, Pure Storage. “It was built for enterprises wondering how we get to AI. It is the first offering to simplify AI at scale, operationalizing things for companies looking to become AI-first organizations.”
AIRI provides the architecture that allows data architects, scientists and business leaders to harness the true power of AI. Burr said that Pure and NVIDIA developed it as the result of requests from joint customers like Montreal-based ElementAI, which makes a platform for companies to build AI solutions, and Zenuity, a joint venture between Volvo and airbag manufacturer Autoliv, which uses AIRI to train its fleet of self-driving vehicles.
“AIRI responds to the evolution of what the next phase of ‘data-driven’ means,” Burr said. “It once meant leveraging Intel spreadsheets. Now it means collecting as much data as I can, dumping it into an unstructured format, and using machines to structure it and generate a result.”
It also goes beyond the traditional premise that software engineers write code.
“With neural networks, based on massive amount of data, machines are beginning to write better and more accurate code than humans, Burr stated. “As a result, Forrester projects that by 2020, $1.2 trillion will be taken from companies that don’t adopt AI to those that do.”
AIRI eliminates the complexity of getting set up on the infrastructure side.
“It allows a plug-and-play onramp to AI,” Burr said. “DGX’s deep learning stack and container-based framework gets you up and running on AI, but it is not fully realized until coupled with the massively parallel storage we provide. That combination facilitates multi-node GTX training. AIRI slashes training time for simulations by nearly 4x. That produces really meaningful advances in how people think about compute.”
AIRI is powered by Pure Storage FlashBlade storage platform, which is architected for modern analytics and AI, and four NVIDIA DGX-1 servers, delivering four Tensor petaflops of performance with Tesla V100 GPUs. These systems are interconnected with Arista 100GbE switches, supporting GPUDirect RDMA for maximum distributed training performance. AIRI also ships with NVIDIA GPU Cloud DL Stack and Pure Storage AIRI Scaling Toolkit.
AIRI also produces huge savings in space and power over the hardware needed to provide the same amount of storage before.
“You would need really large farms – 50 racks of storage and compute – to do this before,” Burr said. “That’s high maintenance. AIRI gives you the equivalent of 50 racks of storage in under 50 inches. It allows multiple data scientists to have the compute and capacity of their own data centre on their own projects.”
AIRI will initially go to market through a select channel of five picked partners, rather than the broader Pure Storage partner base.
“Realistically, there are a limited number of partners who have already invested in this specific practice discipline, Burr indicated. “That is why we will be launching with five partners to get the program started. They have already invested heavily in both Pure and NVIDIA, and focused on this space. These five were at the forefront with both companies.”