Arik Gihon, senior principal engineer of system-on-chip design at Intel, will walk the audience through the design of Lunar Lake, Intel's latest data center chip.
Tiernan Ray of ZDNET
Judging by the well-attended “Hot Chips” chip conference held this week at Stanford University, the science and engineering of building chips dedicated to handling artificial intelligence is hotter than ever.
Now in its 36th year, the Hot Chips show attracts 1,500 attendees, more than half of whom join via an online live feed and the rest at Stanford University's Memorial Auditorium. For decades, the show has been a forum for discussing cutting-edge chips from many vendors, including Intel, AMD and IBM, and companies often use the show to launch new products.
Plus, Linus Torvalds talks about AI, adopting Rust, and why the Linux kernel is “the only thing that matters”
This year's conference received over 100 presentation submissions from around the world. In the end, 24 talks were accepted, roughly as many as could fit into the two-day conference format. There were two tutorial sessions on Sunday, keynotes on Monday and Tuesday, and 13 poster sessions.
The onstage tech talks and poster presentations are very technical and geared towards engineers, and the audience tends to have laptops and multiple screens spread out, just as if they were taking the session in their own office.
Participants tend to set up their laptops and camp out in what seem to be makeshift offices.
Tiernan Ray of ZDNET
Hot Chips 2024 Participants
Tiernan Ray of ZDNET
Monday's morning session featured presentations on Qualcomm's Oryon and Intel's Lunar Lake processors for data centers, drawing a packed audience and raising lots of questions.
There has been a lot of attention in recent years on chips designed to run neural network-style AI more efficiently. At this year's conference, Trevor Cai, OpenAI's head of hardware, gave a keynote address on “Predictable Scaling and Infrastructure.”
Trevor Cai, Infrastructure Engineer at OpenAI, explains the benefits of predictable scaling through increased computing power, which has been a focus for OpenAI since its inception.
Tiernan Ray of ZDNET Tiernan Ray of ZDNET
Cai, who has spent time building OpenAI's computing infrastructure, said ChatGPT is the culmination of the company “spending years and billions of dollars to get better at predicting the next word,” which has led to a set of features like “zero-shot learning.”
“How do we know that will work?” Tsai asked rhetorically. That's because there are “scaling laws” that dictate that power can predictably grow as a “power law” of the math used. Every time the math doubles, precision approaches an “irreducible” entropy, he explained.
Also, what to expect from Meta Connect 2024: A more affordable Quest, AR glasses, and more
“That's why we can invest and build large clusters of computers,” Tsai said. There are “huge headwinds” to continue along the scaling curve, Tsai said. OpenAI will have to tackle some very difficult algorithmic innovations, he said.
When it comes to hardware, “the expenses and energy costs of these large clusters are significant even for the companies that generate the most free cash flow,” Cai said.
The conference continues on Tuesday with presentations from Advanced Micro Devices and startup Cerebras Systems, among others.
