AI chips thrive in Silicon Valley despite Chinese restrictions

Hello and welcome to Protocol Enterprise! Today: why AI chipmakers are optimistic about the future, how the enterprise metaverse could collapse, and why enterprise AI is the new open-source business opportunity.

Oh fries

In a hallway outside a hotel ballroom deep in Silicon Valley, the senior vice president of AI chip startup SambaNova Systems, Marshall Choy, boasted that there are a few moments at the company’s booth, he had signed a $10 million contract for the company’s hardware. Choy displayed what he said was the contract on his phone – too quickly to discern the new client – ​​and walked away.

SambaNova and dozens of other AI startups descended on Santa Clara, California, for the summit of AI hardware. The two-day summit, with a modest exhibit hall and technical discussions, was a significantly less complicated affair than a concurrent investor event at San Francisco’s Palace Hotel hosted by Goldman Sachs.

  • More evidence emerged during the Goldman conference on the volatile potato chip market.
  • An executive at flash and hard drive storage maker Western Digital said hyperscale cloud computing companies were becoming increasingly cautious with their purchases.
  • While China could be a big potential market for companies like Qualcomm, WD and other cloud hardware vendors, economic turmoil has hurt sales there.
  • Nvidia’s Colette Kress echoed WD’s comment about cloud hardware demand in China, describing it as “muted” during her Q&A with a Goldman analyst on Tuesday.

The steady pace of restrictions on various chip-related technologies exported to China — most recently, notices received by Nvidia and AMD regarding advanced AI chips — prompted Goldman Sachs analysts to poll nearly every chip company there about the potential effects on their business.

  • Kress said the company is working with customers to secure licenses or alternative products, though it’s worth noting that the gated A100 chips offer significant performance gains over the previous generation.
  • AMD Data Center Senior Vice President Forrest Norrod said the notification took the company by surprise, but due to recent economic trends AMD has not made China a center of its business. .
  • The largest chip tool maker in the United States, Applied Materials, said most of its business in China was tied to less advanced manufacturing systems, and upcoming export control rules will have minimal impact. on its operations.

In contrast, none of the chip-making startups had received notification letters. from the Commerce Department, among those interviewed by Protocol at the AI ​​Summit. The focus at the top was far removed from geopolitics and was more nerd-focused on silicon and software.

  • One of the most interesting presentations was from Alexis Bjorlin, Meta’s recently hired VP of Infrastructure Engineering, who explained how the company is deploying AI at the mammoth scale needed.
  • Bjorlin, who has a background in optical network infrastructure, said the company operates more than 100 data centers around the world. Beyond 2025, the company plans to build AI training clusters with around 4,000 accelerators.
  • And RISC-V startup SiFive buried some news midway through its Thursday presentation: The company is working with Google to use its X280 cores inside its tensor processing units, or TPUs, which power Google’s AI. .

— Max A. Cherney (E-mail | Twitter)

A MESSAGE FROM THE CNCF

The Cloud Native Computing Foundation’s flagship conference brings together adopters and technologists from leading open source and cloud native communities in Detroit, Michigan, October 24-28, 2022. Register now and join thousands of attendees, including maintainers for the 140 CNCF graduates, incubators and Sandbox projects, virtually or in person.

Register to attend: In person | Virtual

The metaverse has three compartments

Fifty-two percent of executives surveyed this week at Goldman Sachs’ inaugural Communacopia + Technology conference see gaming as the most promising market for the metaverse over the next three years, while 21% think other areas entertainment will thrive, such as concerts and sports. .

Judson Althoff, chief commercial officer of Microsoft, acknowledged the differing opinions on the subject. “Before you all look up to this metaverse… You pull 100 people into a room, you get 100 different answers about what the metaverse is, and will anyone ever get any real value out of it,” a Althoff told the San Francisco conference, according to a transcript of his interview with Goldman Sachs chief executive Kash Rangan.

But Althoff said he sees the metaverse as falling into three categories: there’s the consumer metaverse where Althoff’s avatar could buy Rangan’s avatar a beer, and there will be a monetization aspect. in this area, he said.

Also, “there’s the business metaverse where people will have more engaging and experiential collaboration in the metaverse, and I think there’s an opportunity there to bring people from around the world with different perspectives to collaborate” , did he declare.

Then there’s what Althoff calls the industrial metaverse.

“Think of it as combining sets of technologies, IoT capabilities where you walk in and create a central fabric for any industrial process, any manufacturing environment, any supply chain or logistics scenario,” Althoff said. “You have this central fabric flow, a cloud-based large-scale analytics solution, a big data store. And then you reason about it with machine learning and create what we call digital twins of those environments. and simulate the results.

“If you make or move anything, you create a carbon footprint,” Althoff said. “If I can simulate this for you endlessly in the cloud before I build it and before I move it, I can help you build a better product in a more cost effective way with a lower carbon footprint and water usage. weaker more sustainably than ever before. And we are seeing real progress happening today with companies like Coca-Cola, Unilever, AB InBev, GM, Grupo Bimbo – across all industries – able to model these processes, d save energy, reduce waste, and these solutions pay for themselves.

— Donna Goodison (E-mail | Twitter)

Open source AI will only get you so far

When Meta passed on its popular open source AI framework PyTorch to the nonprofit Linux Foundation on Monday, it reinforced the value the AI ​​industry places on sharing its open source building blocks.

Compared to many other types of technology incorporating proprietary code, AI is a melting pot of shared research and open-source fundamentals that, when tinkered with and customized, help create tools and products of AI.

But there’s a bridge to cross between connecting free AI puzzle pieces and assembling a production system that works at commercial scale. And it’s a toll bridge. (Cue the Python gag).

“[Open-source AI] can make a cool demo; you could create so much cool stuff for free with all the open source stuff out there,” said Davis Sawyer, co-founder and chief product officer of Canada’s Deeplite, which provides software that compresses AI so it’s not too large to work in peripheral devices such as phones or vehicles.

But, he said, “open-source will get you 80 percent. Build a fancy product that goes into a security camera [for instance] — it’s still a significant investment. It’s still going to take millions of dollars even if you’re building on open source.

The difference between a just OK open source-based AI product and a great product depends on the expertise of its builders.

“The secret sauce today is assembly,” said Abhishek Gupta, senior AI manager and expert at the Boston Consulting Group and former machine learning engineer at Microsoft. “The real difference is your ability to experiment and iterate, which comes from the maturity of the underlying software infrastructure, the underlying engineering practices,” he said.

“It’s the distinction between people who have good AI systems and people who have great AI systems.”

-Kate Kaye (E-mail | Twitter)

A MESSAGE FROM THE CNCF

The Cloud Native Computing Foundation’s flagship conference brings together adopters and technologists from leading open source and cloud native communities in Detroit, Michigan, October 24-28, 2022. Register now and join thousands of attendees, including maintainers for the 140 CNCF graduates, incubators and Sandbox projects, virtually or in person.

Register to attend: In person | Virtual

Thanks for reading – see you Monday!

Comments are closed.