OpenAI, Intel, and Qualcomm talk AI compute at legendary Hot Chips conference

The science and engineering of creating chips devoted to processing synthetic intelligence is as vibrant as ever, judging from a well-attended chip convention happening this week at Stanford College referred to as Sizzling Chips.

The Sizzling Chips present, at present in its thirty sixth 12 months, attracts 1,500 attendees, simply over half of whom take part by way of the net reside feed and the remainder at Stanford’s Memorial Auditorium. For many years, the present has been a hotbed for dialogue of essentially the most cutting-edge chips from Intel, AMD, IBM, and lots of different distributors, with corporations typically utilizing the present to unveil new merchandise. 

This 12 months’s convention acquired over 100 submissions for presentation from everywhere in the world. Ultimately, 24 talks had been accepted, about as many as would slot in a two-day convention format. Two tutorial periods befell on Sunday, with a keynote on Monday and Tuesday. There are additionally 13 poster periods. 

The tech talks onstage and the poster shows are extremely technical and oriented towards engineers. The viewers tends to unfold out laptops and a number of screens as if spending the periods of their private workplaces. 

Monday morning’s session, that includes shows from Qualcomm about its Oryon processor for the information heart and Intel’s Lunar Lake processor, drew a packed crowd and elicited loads of viewers questions. 

Lately, an enormous focus has been on chips designed to run neural community types of AI higher. This 12 months’s convention included a keynote by OpenAI’s Trevor Cai, the corporate’s head of {hardware}, about “Predictable scaling and infrastructure.” 

Cai, who has spent his time placing collectively OpenAI’s compute infrastructure, mentioned ChatGPT is the results of the corporate “spending years and billions of {dollars} predicting the following phrase higher.” That led to successive skills similar to “zero-shot studying.”

“How did we all know it might work?” Cai requested rhetorically. As a result of there are “scaling legal guidelines” that present capability can predictably enhance as a “energy regulation” of the compute used. Each time computing is doubled, the accuracy will get near an “irreducible” entropy, he defined. 

“That is what permits us to make investments, to construct large clusters” of computer systems, mentioned Cai. There are “immense headwinds” to persevering with alongside the scaling curve, mentioned Cai. OpenAI must grapple with very difficult algorithm improvements, he mentioned.

For {hardware}, “Greenback and vitality prices of those large clusters grow to be vital even for highest free-cash-flow producing corporations,” mentioned Cai.

The convention continues Tuesday with shows by Superior Micro Gadgets and startup Cerebras Programs, amongst others.