Actionable news
0
All posts from Actionable news
Actionable news in INTC: Intel Corporation,

Intel’s Khosrowshahi: ‘We’re Leading the Way’ in A.I. Processors

Intel's "Intel Nervana Neural Network Processor" for machine learning.

Khosrowshahi is a co-founder, along with Naveen Rao, another Intel executive, of a startup called Nervana Systems. I profiled Nervana in October of 2015 for a Barron’s cover story about how the rise of cloud computing and artificial intelligence was changing the objectives, and the design, of silicon.

After that article, in which it was posited that Intel’s microprocessor business might be in trouble, Rao and Khosrowshahi’s company was bought by Intel, in August of last year, for terms that were not disclosed, but speculated by the New York Times’s Steve Lohr to be north of $400 million.

Rao and Khosrowshahi have both been on the road in support of Intel this year. Rao was at Intel’s Xeon unveiling in July in New York. This time it was Khosrowshahi’s turn.

The first fruits of their labors at Intel will show up in what’s being called an "Intel Nervana Neural Network Processor,” formerly known as Lake Crest, which is expected to be available to some customers by the end of this year.

Rao wrote an essay about the part on October 17th, after Intel CEO Brian Krzanich discussed it at The Wall Street Journal’s “D” conference.

While Nvidia’s (NVDA) GPUs are widely popular for machine learning, and Xilinx (XLNX) FPGAs are becoming more important — and new startups are taking totally different approaches, such as Cerebras Systems, which I wrote about recently — Khosrowshahi sees Intel in a prime position, thanks to innovative design, chip-making prowess, and software tools.

This is the world's first neural network processor,” Khosrowshahi told me. “We are leading the way."

Khosrowshahi describes the the Nervana part as a “distributed linear algebra processor,” with some “sugar” thrown in, in order to “do compounding linear algebra along with some other operations that are important for neural networks."

Khosrowshahi worked extensively with Nvidia GPUs when doing his PhD in neuroscience. And Nervana did a lot of work with them in startup mode, so he and Rao know very well the appeal of that chip platform. A large part of it is "CUDA," the software tools that let one program the GPU. Nvidia's been refining those tools for 11 years, and they now form an important part of Nvidia's advantage in the market.

"New technology is hard to adopt," he concedes. But he says Intel's "MKLDNN" software is kind of a "recreation" of CUDA's "CUDANN," an equivalent software kit, he believes it will go a long way to helping with the learning curve for the Nervana chip.

While both Rao and Khosrowshahi talked with me for two hours in 2015 about changing the “architecture” — the basic functional design — of processors, they have now become convinced of the many advantages of Intel’s silicon expertise as well.

Khosrowshahi waxed enthusiastic during our chat today about the numerous developments “that have been going on here for years,” including “heterogeneous stacks” of logic and memory, and silicon photonics, and carbon nanotubes. In short, he and Rao see Intel’s manufacturing prowess as a key advantage, a superior “substrate,” on top of which they can build the algorithmic and computational designs for neural networks.

"We have the right architecture and the best manufacturer,” says Khosrowshahi.

I asked Khosrowshahi about another avenue of innovation: Domain-specific custom processors.

In particular, the work of David Patterson, a computer scientist with the University of California, Berkeley, who helped Alphabet's (GOOGL) Google to develop the "TPU" processor used for "TensorFlow," a machine learning set of frameworks.

In an article in Barron's in July, I quoted Patterson as saying the big, general purpose processor of old is now going to give way to more domain-specific, reprogrammable processors such as the TPU.

“Instead of the Honda for everyone, we are making these Formula One race cars for some things," as Patterson put it.

Khosrowshahi is respectful of Patterson, who created the "RISC" chip design technology, and of his colleague, John Hennessy, with whom Patterson has published many volumes of a prestigious textbook on chips, "Computer Architecture: A Quantitive Approach."

"Patterson and Hennessy are the founding fathers," he says, "but I would offer a dissenting opinion."

The TPU, says Khosrowshahi, is "actually quite simple, it's a large systolic array, it's not a general purpose processor." There have been times in past when such systolic array approaches to computing can be useful, but he sees limits to it. "Sure you can make the first step with this systolic arrays," but eventually, you'll want to move on to more sophisticated, and powerful approaches, like what he and Rao are building, he contends.

The first Nervana chip is expected to be followed by a part called “Spring Crest,” and then Khosrowshahi says there’ll be a “regular cadence” of new parts after that, on a yearly basis.

Intel shares today are up 5 cents at $40.88.


More