This week, inside a cavernous room in a one-story building in Santa Clara, California, a 6.5-foot-tall machine was spinning behind a white cabinet. These machines made up a new supercomputer that just went live last month.
The supercomputer, unveiled Thursday by Silicon Valley startup Cerebras, is built with the company’s specialized chips designed to power artificial intelligence products. The chip stands out for its size, being about the size of a dinner plate, or 56 times his size for chips commonly used for AI. Each Cerebras chip packs the computing power of hundreds of conventional chips.
Cerebras announced that it has built a supercomputer for AI company G42. G42 said it plans to use supercomputers to develop and enhance AI products for the Middle East.
“What we’re showing here is that there’s an opportunity to build a very large purpose-built AI supercomputer,” said Andrew Feldman, CEO of Cerebras. He added that his startup “want to show the world that this job can be done faster, with less energy and at a lower cost.”
Demand for computing power and AI chips has skyrocketed this year, fueled by the global AI boom. Tech giants such as Microsoft, Meta, and Google, as well as countless start-ups, have been rolling out AI products in recent months after the eerie, human-like prose generated by AI-powered ChatGPT chatbots made headlines. I’m in a hurry.
But manufacturing AI products typically requires a large amount of computing power and specialized chips, so there is a furious demand for more technology. Nvidia, the leading maker of chips that power AI systems, said in May that demand for its product, known as a graphics processing unit (GPU), was so strong that the company’s quarterly sales topped Wall Street. He said it would exceed expectations by more than 50 percent. The prediction sent Nvidia’s market value skyrocketing past $1 trillion.
“For the first time, we are seeing a huge rise in computer requirements” due to AI technology, said Ronendaal, founder of Run:AI, a Tel Aviv startup that helps companies develop AI models. increase. That has created a “massive demand” for specialty chips and companies “are scrambling to secure access” to them, he added.
In order to obtain enough AI chips, some of the big technology companies such as Google, Amazon, Advanced Micro Devices and Intel have developed their own alternative chips. Startups such as Cerebras, Graphcore, Groq and SambaNova are also joining the race, looking to tap into a market that has been dominated by Nvidia.
Chips will play a very important role in AI, potentially changing the balance of power between tech companies and even nation states. For example, the Biden administration recently considered restricting the sale of AI chips to China, and some U.S. officials believe China’s AI capabilities could help the U.S. by bolstering China’s military and security apparatus. He said it could pose a national security threat.
AI supercomputers have been built before by Nvidia and others. But startups rarely create them.
Sunnyvale, Calif.-based Cerebras was founded in 2016 by Feldman and four other engineers with the goal of building hardware that accelerates AI development. Over the years, the company has raised $740 million, including from venture capital firms such as Sam Altman, who heads AI lab OpenAI, and Benchmark. Cerebras is valued at $4.1 billion.
Because the chips used to power AI are typically small, often the size of a postage stamp, hundreds or even thousands of chips are required to handle complex AI models. In 2019, Cerebras unveiled what it claimed to be the largest computer chip ever made, and Feldman said the company’s chip could power AI systems 100 to 1,000 times faster than existing hardware. can be trained.
Abu Dhabi company G42 will start working with Cerebras in 2021. The company used his Cerebras system in April to train the Arabic version of his ChatGPT.
In May, G42 asked Cerebras to build a network of supercomputers around the world. G42 CEO Talal Al-Qaesi said the cutting-edge technology will allow the company to create chatbots and use AI to analyze genomic and preventive health data. rice field.
But demand for GPUs was so high that it was difficult to get enough to build a supercomputer. Al-Qaisi said Cerebras’ technology is available and cost-effective. So Cerebras used its chips to build a supercomputer for the G42 in just 10 days, Feldman said.
“The timescale has been significantly shortened,” Al-Qaesi said.
Cerebras says it plans to build two more supercomputers for the G42 over the next year, one in Texas and one in North Carolina, and then distribute six more around the world. We call this network the Condor Galaxy.
Chris Manning, a computer scientist at Stanford University whose research focuses on AI, said it would still be difficult for the startup to compete with Nvidia, saying that the people building the AI models are the ones who want Nvidia’s AI. That’s because he’s used to using software that runs on chips, he said. .
Other startups have also tried to enter the AI chip market, but many have “virtually failed,” Dr. Manning said.
But Feldman said there is hope. Many AI businesses don’t want to be tied to Nvidia alone, he said, and there is global demand for other powerful chips like those from Cerebras.
“I hope this moves AI forward,” he said.
