The Supercomputing 2017 (SC17) conference has been in full swing this week, with major announcements from a number of players in the HPC industry. Nvidia has been no exception — the company has feathered its cap with announcements that it now has major co-processors in 87 of the TOP500 supercomputers worldwide, as well as six of the ten most efficient supercomputers in the world (listed as the GREEN500). Now, Nvidia is putting its hard work on Volta back into its own hardware, with a massive upgrade to the SaturnV supercomputer it built for its own use.
A year ago, the SaturnV was the most power-efficient system in the TOP500, with a rated capability of 9.46GFLOPS/watt. Today, that same rating only qualifies the system for #14. That’s respectable, to be sure, but Volta does appear to deliver a fair kick — though we can’t estimate how much. The new SaturnV machine is a fraction the size of the old model, with 22,440 cores versus 60,512 and 16.5TB of memory as opposed to 62TB.
However, the TOP500 comparison doesn’t really capture the scope of the upgrade. It’s absolutely true that the new SaturnV iteration has fewer cores and less total RAM, but it packs far more GPUs. The 2016 model had 125 nodes with eight GPUs per node, for a total of 1,000 GPUs and 3.3PFLOPS of processing power. The SaturnV 2017 Edition has 660 nodes with eight GPUs per node, for a total of 5,280 Volta GPUs.
The building block of the system, in both 2016 and 2017, is the Nvidia DGX-1, a supercomputer-in-a-box that contains eight Volta GPUs per chassis. Nvidia also engineered Volta for supercomputing and AI workloads and has made performance in those areas a critical component of the product’s marketing and positioning. That’s in-line with what we’ve seen from multiple other companies over the past year, whether we’re talking about AMD, Fujitsu, Google, or Intel. Everyone and their kid sister has gone racing towards the magic of AI and deep learning, and multiple hardware companies are working on their own custom silicon to drive these new experiences (or create them in the first place).
It’s going to be genuinely interesting to see how this shakes out. High-end AI battles between rival semiconductor companies might not seem to have much impact on the consumer market, but the technologies these companies are trying to develop — better AI, self-driving cars, medical advances, and many others — could revolutionize human life on Earth just as much as the first PCs did. Nvidia knows that early movers have the best shot at being part of a market long-term, which is why they’re pushing so hard to stay on top of this fast-moving space.
Let’s block ads! (Why?)