NVIDIA’s CEO called 2015 a cornerstone year for A.I, thanks to deep learning, and doubles down on powering deep-learning computing data centers with the NVIDIA DGX-1, a server designed from the ground up to accelerate these types of workloads. NVIDIA says that the DGX-1 is powerful enough to handle the load of 256 regular servers, thus making it a “datacenter in a box”.
Although comparisons like that are often designed to give people a catchy high-level, contrasted view, there is a lot of truth in it, if you’re looking at building a deep-learning datacenter. In the past, it’s been proven that GPU-powered servers could dramatically increase the compute density, replacing thousands of servers with hundreds.
The DGX-1 server pushes the same concept to the extreme. At its core, it uses 8 Tesla P100 chips that are 2X faster than the previous generation. Secondly, its memory bandwidth has been much improved, thanks to NVIDIA’s NVLink memory bus technology. Data transfer rates are often a bottleneck for any computing application. “It’s about how fast you can move data” has been true for servers, but also video games and any other high-performance app.
With 8 Tesla P100 chip that can dissipate 300W (thermal design point) each, I would be curious to see what NVIDIA’s cooling design looks like, but the box offers enough room for many solutions, so this should not be a problem.
At $129,000, it seems like a crazy price, but relative to its compute power and compute density, it may be very attractive to companies building deep-learning data centers (and other High-Performance Computing / HPC applications), and there’s no shortage of those right now. It’s fair to assume that demand will be brisk for this product.