TORONTO — Cerebras Systems, a Silicon Valley semiconductor startup, is doubling down on Toronto’s talent pool as it seeks a bigger share of the hot global market for artificial intelligence infrastructure.  

The firm, founded in Sunnyvale, Calif., in 2016, makes machines that supply the processing power customers need to train and run machine-learning (ML) models. “We build AI supercomputers from the ground up,” said Andrew Feldman, CEO and co-founder of Cerebras, in an interview with The Logic.

The company says its novel processor outperforms market-leader Nvidia’s graphics processing unit, by packing and connecting more AI cores and memory on a single, large wafer. “We build the biggest chip in the world,” Feldman emphasized, “not by a little bit.” Cerebras designs the hardware that links, powers and cool its chips and other components. It also develops software that strings the systems together. 

Clients like the Mayo Clinic, GlaxoSmithKline and Total Energies have bought clusters of the startup’s machines for their own data centres, or run their applications on Cerebras’s cloud service.

The company’s Toronto teams are responsible for some of its innovations. In March 2020, the firm opened its office in the city, where it now employs 100 workers. (Multinational-attraction agency Toronto Global helped facilitate the move). It’s looking to double that headcount over the next 12 to 18 months, Feldman said. The firm has 411 employees globally, with additional offices in San Diego and Bangalore.

Global demand for AI compute is booming as businesses develop and deploy new applications, particularly generative ones based on large language models. Nvidia has been the biggest beneficiary, with its sales and stock price surging to make it one of the world’s most valuable companies.

Cerebras is one of several chip startups and stalwarts wrestling for a share of all the new revenue AI compute could generate. Last week, The Information reported that the firm had confidentially filed for an initial public offering. Feldman declined to comment on a possible listing. 

Former Intel executive Nish Sinnadurai, vice-president of software engineering, runs Cerebras’s Canadian operations. Its Toronto teams work on core components, including the compiler, which controls the hardware and ensures it runs models as fast and efficiently as possible. “It was a very good place to find people who worked on very hard software problems,” said Feldman. The firm has also hired ML engineers and chip designers in the Toronto office. 

The Toronto area has long housed research and development hubs for U.S. chip giants, often built around the Canadian upstarts they acquired. “It used to be that Intel and Altera had a monopoly on the most talented folks coming out of the universities,” noted Sinnadurai. But Cerebras has attracted new graduates and experienced professionals in Toronto by promising they’ll work on the firm’s key technologies, not a multinational’s side-project.  

Cerebras is “really pushing the boundary, technologically,” said Vaughn Betz, a University of Toronto researcher who will join the chipmaker as a distinguished visiting professor for his sabbatical year starting in September. He cited its novel chip architecture, manufacturing methodology and software systems.

Feldman sees commercial and funding opportunities in Canada; he’s had discussions with Canadian institutional investors, including pension funds. Cerebras has raised US$723 million to date, according to PitchBook data, with blue-chip backers including Benchmark, Coatue Management and Eclipse Ventures. 

Some newer chip players—like Toronto’s Untether and Tenstorrent, and Groq, another Silicon Valley startup—focus on inference, the stage when a model produces outputs like text, images or autonomous-driving signals. 

Feldman says Cerebras’s systems are competitive for both inference and the earlier, processing-intensive training stage. The dual capability allows it to rival the likes of AMD and Nvidia for buyers looking to fill facilities with AI-optimized machines. “If you’re in a data centre, there’ll be a real advantage to having the ability to do both,” said Feldman. 

Unlike some competitors, Feldman said his firm has focused on the basic mathematics of AI rather than particular forms like the transformers that underpin LLMs. “Hardware has long lead times, and the models are changing quickly,” he noted, but with its approach “we’re setting high watermarks, across the board.”

Governments have pledged billions of dollars to secure compute capacity within their own borders, to keep hold of sensitive citizen data or ensure access for domestic innovators and users. Abu Dhabi-based G42, the emirate’s AI play, is building infrastructure in the region using machines from Cerebras, in which it has invested; the two firms have also launched three supercomputing clusters in the U.S. 

Feldman listed several countries in which Cerebras is exploring private data centres or government sovereign compute deployments, including Azerbaijan, Iceland, India, Finland, Kazakhstan, the U.K., Saudi Arabia and Turkmenistan. G42 will partner on some projects. 

In addition to competing with other chipmakers to sell in those places, Cerebras will also be competing with them for talent, including in Toronto. There are local players Untether and Tenstorrent. Intel and Altera—which it recently spun out—both still have major labs in the city, as do AMD and Nvidia. 

And other newer U.S. entrants to the semiconductor business are also setting up in Canada’s largest city. The Amazon Web Services team that develops the software development kits for the cloud service’s custom chips is based in Toronto, said Matt Woods, vice-president for AI products, in an interview with The Logic. Groq’s compiler team is also run out of Canada. 

Toronto has plenty of workers with expertise in computer-aided design tools, spatial computing and other fields that marry software and hardware, according to Betz, who sold a startup to Altera and has worked with Intel. “Companies are coming here to take advantage of that talent base,” he said.