Buzzwords like "AI" and "machine learning" are thrown around like candy these days, but the use cases for those technologies are growing at a rapid pace. Microsoft and Nvidia see burgeoning demand for cloud-based artificial intelligence in fields like healthcare, voice recognition, molecular simulations, and self-driving vehicles. To accelerate the deployments of AI-oriented technologies, the two companies have been working together on the HGX-1, a "hyperscale" GPU accelerator chassis.
The HGX-1 chassis is powered by eight of Nvidia's Tesla P100 graphics cards. To offer customers considerable flexibility and scalability in the configuration of their systems, the HGX-1 uses a switching design based on Nvidia's NVLink technology that allows a CPU to connect dynamically to as many GPUs as necessary. With four HGX-1 units, for example, users can tap into the power of 32 GPUs total.
So how much performance does this dizzying array of computing power deliver? Nvidia claims that compared with "legacy" CPU-based servers, the HGX-1 offers 100x faster performance in common deep-learning tasks. While the HGX-1 chassis certainly won't come cheap, Nvidia estimates that its customers will be able to conduct AI training at one-fifth the cost of current methods, and AI inferencing at one-tenth the cost.
The hardware is part of Microsoft's ongoing Project Olympus initiative, its contribution to the Open Compute project. Nvidia and Microsoft hope that sharing the project with this open-source hardware development consortium will make the design easier for enterprises to purchase and deploy.