Inside a large room this week in a one-story building in Santa Clara, Calif., six-and-a-half-foot-tall machines spin behind white cabinets. The machines make up a new supercomputer that became operational last month.
The supercomputer, announced Thursday by Cerebras, a Silicon Valley start-up, is built using the company’s special chips, designed to power artificial intelligence products. The chips are remarkable for their size — about the size of a dinner plate, or 56 times larger than a chip typically used for AI Each Cerebras chip contains the computing power of hundreds of traditional chips.
Cerebras says it built the supercomputer for G42, an AI company. The G42 said it planned to use the supercomputer to create and power AI products for the Middle East.
“What we’re showing here is that there is an opportunity to build a massive, dedicated AI supercomputer,” said Andrew Feldman, Cerebras’ chief executive. He added that his start-up wants to “show the world that this work can be done faster, it can be done with less energy, it can be done at a lower cost.”
Demand for computing power and AI chips has soared this year, fueled by the global AI boom. Tech giants like Microsoft, Meta and Google, as well as a host of start-ups, have rushed to launch AI products in recent months after the AI-powered ChatGPT chatbot went viral for the eerily human-like prose it could generate.
But creating AI products typically requires large amounts of computing power and specialized chips, leading to a frantic search for more of those technologies. In May, Nvidia, the leading maker of chips used to power AI systems, said that appetite for its products – known as graphics processing units, or GPUs – was so strong that its quarterly sales would be more than 50 percent above Wall Street estimates. The forecast sent Nvidia’s market value soaring above $1 trillion.
“For the first time, we’re seeing a huge increase in computer requirements” due to AI technologies, said Ronen Dar, a founder of Run:AI, a Tel Aviv start-up that helps companies develop AI models. That “created a huge demand” for specialized chips, he added, and companies “rushed to secure access” to them.
To get their hands on enough AI chips, some of the biggest tech companies – including Google, Amazon, Advanced Micro Devices and Intel – have developed their own alternatives. Start-ups such as Cerebras, Graphcore, Groq and SambaNova have also joined the race, aiming to break into the market dominated by Nvidia.
Chips are set to play such an important role in AI that they could change the balance of power among tech companies and even nations. The Biden administration, for one, has recently weighed restrictions on the sale of AI chips to China, with some American officials saying China’s AI capabilities could pose a national security threat to the United States by enhancing Beijing’s military and security apparatus.
AI supercomputers have been built before, including Nvidia. But it’s rare for start-ups to create them.
Cerebras, based in Sunnyvale, Calif., was founded in 2016 by Mr. Feldman and four other engineers, with the goal of building hardware that accelerates the development of AI. Over the years, the company has raised $740 million, including from Sam Altman, who heads AI lab OpenAI, and venture capital firms like Benchmark. Cerebras is valued at $4.1 billion.
Because the chips typically used to power AI are small — often the size of a postage stamp — it takes hundreds or even thousands of them to process a complex AI model. In 2019, Cerebras took the wraps off what it says is the largest computer chip ever made, and Mr. Feldman said its chips can train AI systems between 100 and 1,000 times faster than current hardware.
G42, the Abu Dhabi company, began working with Cerebras in 2021. It used the Cerebras system in April to train the Arabic version of ChatGPT.
In May, the G42 asked Cerebras to build a network of supercomputers in different parts of the world. Talal Al Kaissi, the chief executive of G42, said that the new technology will allow his company to create chatbots and use AI to analyze genomic data and preventive care.
But the demand for GPUs is so high that it’s hard to get enough to build a supercomputer. Cerebras’ technology is both available and cost-effective, said Mr. Al Kaissi. So Cerebras used its chips to build a supercomputer for the G42 in just 10 days, Mr. Feldman said.
“The time scale has been reduced dramatically,” said Mr. Al Kaissi.
Next year, Cerebras said, it plans to build two more supercomputers for the G42 – one in Texas and one in North Carolina – and, after that, six others distributed around the world. It calls this network the Condor Galaxy.
Start-ups are however likely to have difficulty competing against Nvidia, said Chris Manning, a computer scientist at Stanford whose research focuses on AI. That’s because people who develop AI models are used to using software that works with Nvidia’s AI chips, he said.
Other start-ups have also tried to enter the AI chips market, but many have “effectively failed,” said Dr. Manning.
But Mr. Feldman said he was hopeful. Many AI businesses don’t want to be locked into just Nvidia, he said, and there is global demand for other powerful chips like those from Cerebras.
“We hope it will advance AI,” he said.