Nvidia to Intel: 'Which HPC chip brain will win? Let the people decide'
Easy to say when you have the lion's share of the market already, Little Green
SC13 Nvidia is of the opinion that in the tussle with Intel as to whether Nvidia's Tesla GPU accelerators or Chipzilla's Xeon Phi many-core CPU coprocessors will dominate the supercomputing market, the HPC community has already voted with its checkbooks.
"This community continues to invest in GPU computing – GPU-based accelerated computing – and that actually is the proof in the pudding," Sumit Gupta, general manager of Nvidia's HPC-focused Tesla biz, told The Register when we sat down with him on Wednesday morning at the SC13 supercomputing conference in Denver, Colorado.
On Tuesday, Intel technical computing hardware headman Rajeeb Hazra told a reporters' roundtable that GPU-accelerated computing was essentially just a phase in the development of accelerated high-performance computing. According to Hazra, his next-generation Xeon Phi, "Knights Landing", will usher in an era of lower latency and higher efficiency because it won't be necessary to offload data to GPU accelerators, but instead run HPC workloads directly on this many-core CPU.
"It doesn't matter what I or Raj say," Gupta told us. "It only matters what the HPC community does, and they continue to invest in GPU-accelerated computing. They do not invest in Xeon Phi."
And there are plenty of performance reasons why the HPCers are increasingly embracing GPU-accelerated computing, Gupta believes. For one, he takes issues with the idea of running both the operating system and compute tasks on the same silicon, as will Intel's Knights Landing.
"If you look at any supercomputer today," he said, "any large system, they have what they call host nodes and compute nodes. They actually separate out the operating system from the compute. They don't want it to be together, because if you do the management tasks on the same device, you're going to see what's known as 'jitter' – the computing device will have to, every once in awhile, service other requests. Everything else waits."
Gupta was quick to add that such management overhead wouldn't necessarily be a show-stopper. "I'm not saying this is a bad thing for everyone," he said. "Some applications may get the performance out of it, but I think a large majority won't."
He also said that – as of today, at least – Nvidia's Tesla has a significant performance lead over Intel's Xeon Phi. "Just from a performance-to-performance basis, today the current "Knights Corner" [Xeon Phi] products are almost half the performance of a Tesla in term of real applications," he said, referring to the Tesla K20X, which was just replaced by the more-potent Tesla K40. "Real applications are two times faster on a Tesla. I don't see Intel catching up to that."
'You can't displace someone by being 10 per cent better'
And even if Knights Landing closes that gap, Gupta believes that Tesla's growing number of high-performance installations will be a formidable barrier to Intel taking a bite out of Nvidia. "In this entire market – in the enterprise market and the HPC market – you can't displace someone by being 10 per cent better. You have to be 2X better," he told us.
Gupta also said that by the time Knights Landing ships in "2015, 2016," there will be "hundreds of thousands" of developers coding for Tesla systems using Nvidia's CUDA language. We reminded him that there are tens of millions of x86 developers globally who could code for Knights Landing, but that Intel developer lead over CUDA didn't impress him.
"They have to completely change their application to use Knights Landing," he said. "x86 has nothing to do with the Knights family. The fact that they use an x86 core gives them no advantage, because no one writes in assembly. Everyone writes at a high level, and you have to program a multicore device."
What it all comes down to is market acceptance by customers who have been briefed on the future plans of both Nvidia and Intel. "Everyone that's using GPUs today knows our roadmap; they know Intel's roadmap. If they thought our roadmap was bad, they'd stop developing," he said.
"IBM saw both our roadmaps, and they chose to come work with us."
That said, Gupta welcomes Intel – which he called a "formidable competitor" – into the HPC accelerator/coprocessor space. "I'm happy to have a competitor," he said. "I'm happy Intel is playing in this market."
When we asked what the future will bring in the HPC race, Gupta brought it all back to customers. "I hate talking about the futures of all these kinds of things," he said, "because I can speculate anything I want and make up anything I want. All that matters is what customers are doing today."
And tomorrow, we reckon. And from where we sit, Gupta is spot-on about one incontrovertible truth: when it comes to market dominance, all that matters is which company's name is most often written after "Pay to the order of" on customers' checks. ®