The company kicked off its SC activities with a press briefing on Monday (Nov. 14), during which CEO Jen-Hsun Huang characterized 2016 as a tipping point for the GPU computing approach popularized by Nvidia for over a decade.
Not surprisingly, Huang’s main message was that the GPU computing era has arrived. Throughout the hour-long talk, Huang would revisit the theme of deep learning as both a supercomputing problem and a supercomputing opportunity.
“We believe that supercomputers ought to be designed as AI supercomputers – meaning it has to be good at both computational science as well as data science – that building a machine that’s only good at data science doesn’t make sense and building a supercomputer that’s only good at computational science doesn’t make sense,” he said.
“On the one hand, deep learning requires an enormous amount of data throughput processing – this way of developing software where the computers write software themselves inspired by a lot of data processing behind it is a very important approach to computing but it also has the wonderful opportunity to benefit supercomputing as well, solving problems for science that hasn’t been possible before today,” said Huang.
While Nvidia is enabling parallel computing via thousands of CUDA cores combined with the CUDA programing framework, the CEO emphasized the necessity of a performant central processing unit. “Almost everything we do we start with a strong CPU,” said Huang. “We still believe in Amdahl’s law; we believe that code has a lot of single threaded parts to it and this is an area that we want to continue to be good at.”
The two servers currently shipping with the NVLink P100 GPU – Nvidia’s DGX-1 server and IBM’s Minsky platform – speak to this goal. The DGX-1 connects eight NVLink’d Pascal P100s to two 20-core Intel Xeon E5-2698 v4 chips. The IBM Minsky server leverages two Power8 CPUs and four P100 GPUs connected by NVlink up to the CPUs.
Nvidia’s 124-node supercomputer, SaturnV plays a crucial role in Nvidia’s plans to usher in AI supercomputing. The machine debuted on the 48th TOP500 list at number 28 with 3.3 petaflops Linpack (4.9 petaflops peak). Even more impressively, it nabbed the number one spot on the Green500 list achieving more than 8.17 gigaflops/watt. That’s a 42 percent improvement from the 6.67 gigaflops/watt delivered by the most efficient machine on the previous TOP500 list. Extrapolating to exascale gives us 105.7 MW. If we go with a semi-“relaxed” exascale power allowance of 30 MW (the original DARPA target was 20 MW), this is less than one-fourth the planned power consumption of US exascale systems. Three years ago, the extrapolated delta was over a 7X.
Huang noted that when you apply deep learning FLOPS math – aka 16-bit floating point operations as opposed to the HPC norm of 64-bit FLOPS, exascale is not far away at all.
The [IBM/Nvidia] CORAL machines are on track for 2018 with 300 petaflops peak FP64, which comes out to 1,200 peak FP16, Huang pointed out. “For AI, FP16 is fine, now in some areas we need FP32, we need variable precision, but that’s the point,” he said. “I think CORAL is going to be the world’s fastest AI supercomputer [and] I think that we didn’t know it then but I believe that we are building an exascale machine already.”
SaturnV is organized into five 3U boxes per rack, with 15 kilowatt of power on each rack and some 25 racks total. While the press photo of SaturnV indicates 10 servers per rack, this is not reflective of what’s inside. “We could not put that many in ours,” said Capps. “We put this in a datacenter which is not HPC. It was an IT datacenter originally.”