NVIDIA seems to be mounting a vigorous effort to dethrone the CPU as the leader of the processor pack for HPC and demanding datacenter workloads. That’s a tall order. Introduction at ISC 2016 of a PCIe-based version of is new Tesla P100 card is one element in the strategy. It should ease the upgrade path for many. But really, NVIDIA’s mantra has been shifting since introduction of the Pascal chip and DGX-1 platform last April. NVIDIA argues that better performance scaling on GPUs (Intel would dispute this), the mainstreaming of AI/deep learning approaches in scientific computing, and the large base of CUDA developers and code are combining to lift GPUs from a supportive to preeminent role.
“If you look at real applications, they tend not to scale well as you add CPU only nodes,” argues Marc Hamilton, vice president, solutions architecture and engineering at NVIDIA, citing benchmark data for Amber (molecular dynamics) running on Comet at the San Diego Supercomputing Center. “After 30 or 40 CPUs, no matter how many CPUs you add the curve doesn’t scale. Both of these charts argue for having a having a faster GPU and a more powerful compute node.” (See slide below.)
GPUs (and acceleration generally) have been transforming HPC for some time. Now the paradigm is shifting in the datacenter as well, says Hamilton who notes pointedly, “The new P100 PCIe card is targeting general purpose high performance computing. Our second ISC theme is AI is next big thing for HPC. There’s no disputing that from the consumer web side for apps [and] they use GPUs for network training. [Moreover] every major supercomputing center of the past six months is really using AI to address scientific computing that just couldn’t be addressed before.”
Specs for the new Tesla P100 GPU accelerator for PCIe include:
- 4.7 teraflops double-precision performance, 9.3 teraflops single-precision performance and 18.7 teraflops half-precision performance with NVIDIA GPU Boost technology.
- Support for PCIe Gen 3 interconnect (32GB/sec bi-directional bandwidth).
- Enhanced programmability with Page Migration Engine and unified memory.
- ECC protection for increased reliability.
- Server-optimized for highest data center throughput and reliability.
- Available in two configurations: 16GB of CoWoS HBM2 stacked memory, delivering 720GB/sec of memory bandwidth; 12GB of CoWoS HBM2 stacked memory, delivering 540GB/sec of memory bandwidth.
“P100 for PCIe-based servers is basically the same as the NVLink-based P100 card but packaged in a standard PCIe form with a standard PCIe power footprint (300w for NVLink version versus 250w for PCIe),” reports Hamilton. “It is the same chip, the same module. We still use the HBM2 memory, same size memory, 16G, full 720GBS bandwidth to the GPU, and it has support for the page migration engine.”
Hamilton says a single Tesla P100-powered server delivers higher performance than 50 CPU-only server nodes when running the AMBER molecular dynamics code, and is faster than 32 CPU-only nodes when running the VASP material science application. Later this year, Tesla P100 accelerators for PCIe will power an upgraded version of Europe’s fastest supercomputer, the Piz Daint system at the Swiss National Supercomputing Center in Lugano, Switzerland (see Europe’s Fastest Computer to Get Pascal GPU Upgrade).
Just as interesting as the P100 cards is the DGX-1 system marketed as the world’s first GPU-based deep learning supercomputer. Hamilton says NVIDIA has customers buying multiple units and that bundled DGX-1 clusters would quickly crack the TOP500 list.
When first launched, “We weren’t shipping [but] did the math for the ISC list expectations and estimated with 10-to-12 DGX-1s in a single rack you could probably get on the TOP500 list. We always simulate out what our expectations are for the new lists and for November we think on the order of 15-16 DGX-1 boxes should get you on the TOP500 list and we certainly have customers that are looking at deployment of that size and larger,” Hamilton said.
Perhaps we’ll see the emergence of a new TOP500 ‘entrant profile’ that is GPU-only or GPU-mostly. Roughly 100 of the TOP500 are accelerator-assisted machines, most by NVIDIA GPUs.
Hamilton emphasizes NVIDIA sees a broad opportunity for more pervasive penetration by GPU-only/mostly systems throughout the enterprise, especially as deep learning computing techniques are applied to a wider swath of scientific and enterprise workloads. Big Internet service providers, of course, make heavy use of GPU uses for deep learning et al. “Deep learning is a new style of computing and requires boxes like the DGX-1. We’ve seen continued interest from not only the likes of HPE, Dell, IBM, and Cray on NVLink-based designs but also from smaller OEMs as well,” says Hamilton. The latter, presumably, would more be inclined to use new P100 PCIe card.
Critics say GPU systems are more expensive. Hamilton agrees, if the metric is simply server hardware cost. “I think there’s some misconceptions about GPU systems. Of course a CPU server is more expensive because you start with a regular server and add a GPU. However the GPU-enabled datacenter running the same workload is going to be lower cost because not only do you have fewer nodes, but also you reduce all the other infrastructure at a typical supercomputer center. Turns out only 39 percent of the compute budget is used for servers; the rest of it is racking, cabling, networking, power and cooling, etc. One of the goals we’re trying to do is address that 61 percent of the budget,” he says.
The market will eventually decide, as always, but NVIDIA seems to be carefully lining up multiple arguments for its case – not least the need. NVIDIA notes that NSF was oversubscribed by 200 billion core hours, meaning rejected cycles, because its capacity was full and that capacity is heavily CPU-based.
“Accelerated computing is the only path forward to keep up with researchers’ insatiable demand for HPC and AI supercomputing,” said Ian Buck, vice president of accelerated computing at NVIDIA in the company’s official release announcement. “Deploying CPU-only systems to meet this demand would require large numbers of commodity compute nodes, leading to substantially increased costs without proportional performance gains. Dramatically scaling performance with fewer, more powerful Tesla P100-powered nodes puts more dollars into computing instead of vast infrastructure overhead.”