If you want to get under Diane Bryant’s skin these days, just ask her about GPUs.
The head of Intel’s powerful data center group was at Computex in Taipei this week, in part to explain how the company’s latest Xeon Phi processor is a good fit for machine learning.
Machine learning is the process by which companies like Google and Facebook train software to get better at performing AI tasks including computer vision and understanding natural language. It’s key to improving all kinds of online services: Google said recently that it’s rethinking everything it does around machine learning.
It requires a massive amount of computing power, and Bryant says the 72 cores and strong floating point performance of Intel’s new ‘Knight’s Landing’ Xeon Phi, released six months ago, give it an excellent performance-per-watt-per-dollar ratio for training machine learning algorithms.
“It’s a big opportunity, and there will be a hockey stick where every business will be using machine learning,” she said in an interview.
The challenge for Intel is that the processors most widely used for machine learning today are GPUs like those from Nvidia and AMD.
“I’m not aware that any of the Super Seven have been using Xeon Phi to train their neural networks,” said industry analyst Patrick Moorhead, of Moor Insights and Strategy, referring to the biggest customers driving machine learning – Google, Facebook, Amazon, Microsoft, Alibaba, Baidu and Tencent.
Bryant, who is very affable, grew mildly exasperated when asked how Intel can compete in this market without a GPU. The general purpose GPU, or GPGPU, is just another type of accelerator, she said, and not one that’s uniquely suited to machine learning.
“We refer to Knights Landing as a coprocessor, but it’s an accelerator for floating point operations, and that’s what a GPGPU is as well,” she said.
She concedes that Nvidia gained an early lead in the market for accelerated HPC workloads when it positioned its GPUs for that task several years ago. But since the release of the first Xeon Phi in 2014, she says, Intel now has 33 percent of the market for HPC workloads that use a floating point accelerator.
“So we’ve won share against Nvidia, and we’ll continue to win share,” she said.
Intel’s share of the machine learning business may be much smaller, but Bryant is quick to note that the market is still young.
“Less than 1 percent of all the servers that shipped last year were applied to machine learning, so to hear [Nvidia is] beating us in a market that barely exists yet makes me a little crazy,” she says.
Still, 1 percent of the worldwide server market is not trivial, and Intel will continue to evolve Xeon Phi to make it better at machine learning tasks.
It’s not without customers in the area, though it can’t point to household names. Bryant mentioned Viscovery, which is using Knights Landing to train algorithms for video search.
There are two aspects to machine learning, she notes – training the algorithmic models, which is the most compute intensive part, and applying those models to the real world in front-end applications, often called inferencing.
Intel’s FPGAs, acquired from its Altera acquisition, coupled with its regular Xeon processors, are well suited to the inferencing part, Bryant says, so Intel has both sides of the equation covered.
Still, it may have a hard time displacing GPUs at the hyperscale companies – not to mention Google’s TPU, or Tensor Processing Unit, a chip that company built specifically for machine learning.
Nvidia’s GPUs are harder for programmers to work with, Moorhead said, which could work in Intel’s favor, especially as regular businesses start to adopt machine learning. And Knights Landing is “self-booting,” which means customers don’t need to pair it with a regular Xeon to boot an OS.
But Intel’s newest Xeon Phi has a floating point performance of about 3 teraflops, Moorhead said, compared to more than 5 teraflops for Nvidia’s new GP100.
“You could beef up the floating point on Knights Landing and have something that looks like a GPU, but that’s not what it is right now,” he said.
Still, Intel is persistent, and it’s determined to succeed. “We’ll continue to advance the product line, and we will continue to take share,” Bryant said.