Intel’s Neural Net Processor – Too Little, Too Late?

Article By : Dylan McGrath, EE Times

"Intel is late to the party and shooting at a moving target," says analyst

SAN FRANCISCO — Intel disclosed plans to release late next year its first neural network processor intended for commercial use, but some analysts wonder if it will be too late for Intel to make up ground lost in AI to Nvidia’s Volta GPU architecture.  

The Intel chip, codenamed Spring Crest, will offer three to four times the training performance of Intel’s first-generation neural network processor, Lake Crest.

Spring Crest — or the Intel Nervana NNP-L1000 as it is formally known — was the most significant of a number of innovations described by Naveen Rao, vice president and general manager of Intel’s AI Products Group, Tuesday (May 23) at Intel’s first-ever AI developer event here.

Kevin Krewell, a principal analyst at Tirias Research, said that even with three to four times the performance of Lake Crest, Spring Crest would still have only a slight performance advantage over Nvidia’s V100 data center GPUs, which are already available.

“It may take until 2020 or 2021 for Intel to potentially deliver a superior neural net processor,” Krewell said.

“Intel is late to the party and shooting at a moving target,” said Jim McGregor, founder and principal analyst at Tirias. “You can bet that the Nvidia solution will be well beyond that point by that time. It kinda sounds like Intel and GPUs all over again.”

Naveen Rao speaks at Intel's AI DevCon event in San Francisco Tuesday. Credit: Intel

Naveen Rao speaks at Intel’s AI DevCon event in San Francisco Tuesday.
Credit: Intel

Rao and other Intel executives at the AI event sought to portray Nvidia as a one-trick pony, repeatedly extolling the virtues of Intel’s portfolio of chips for AI — including neural net processors, Xeon processors and Stratix FPGAs — as a comprehensive suite of solutions for the rapidly evolving AI landscape.

“It’s not a one size fits all for AI silicon,” Rao said. “You really need application-specific solutions. It’s going to take a combination of architectures.”

The semiconductor industry is going full tilt to bring to market silicon to support the creation of deep neural networks (DNNs). In addition to Intel — which jump-started its AI efforts with the $400 million acquisition of Rao’s Nervana in 2016 — established companies including Nvidia, IBM and Google and a host of startups are pushing out chips designed and optimized to train DNNs, with many slated to be available this year.

Lake Crest, which began at Nervana prior to the acquisition by Intel, is shipping to customers now, “not a competitive processor at this point,” according to Krewell. “It will instead become a software development platform.”

— Dylan McGrath is the editor-in-chief of EE Times.

Subscribe to Newsletter

Leave a comment