On Thursday, Cerebras Systems said that it signed an approximately $100 million deal to supply the first of three artificial intelligence (AI) supercomputers to the United Arab Emirates-based technology group G42. It said Cerebras will build the first of its Condor Galaxy systems in Santa Clara, California, and set up two more in the U.S. in early 2024. G42, which has nine operating companies that include datacenter and cloud services businesses, will use the supercomputers to accelerate AI development across several sectors, including health care and energy.
The company said that the first machine, CG-1, is powered by Cerebras’ CS-2 chip, which features 54 million AI-optimized cores. The system is designed to train large language models and perform deep learning tasks, such as image recognition. It can handle four exaFLOPs of performance, the equivalent of running 500,000 desktop computers for a second, and consumes about half as much power per core as a typical GPU cluster, Cerebras CEO Andrew Feldman said.
Each CS-2 chip is dinner-plate sized and houses 850,000 cores on a single piece of silicon with the equivalent of 2.6 trillion transistors. It also includes a massive memory module on the same piece of silicon that slashes the time needed to shuttle data between the processors and main memory. Feldman said that’s a big part of why the system can scale more efficiently than others.
He added that unlike other AI chips, which are built on a separate piece of silicon, the CS-2s are linked together to work as one system. He said That the system can cut model training times by two orders of magnitude compared with today’s best systems.
He said the new system could also train models of up to 120 trillion parameters, which is twice as large as current brain-scale neural networks. That’s enough to train the world’s most advanced AI systems to process natural language and learn how to drive cars, and perform medical procedures.
Cerebras is among several firms accelerating AI machines’ processing power as the industry grapples with the slowdown in chip sales caused by China’s economic woes. The industry has also been shaken by concerns that the United States could ban Chinese tech firms such as TikTok, the social media phenom whose popularity has surged since it started to allow people in the country to share political content.
