Cerebras Systems has introduced the latest version of its wafer-sized semiconductor.
The Wafer Scale Engine 3 boasts four trillion transistors and 900,000 'AI cores,' alongside 44GB of on-chip SRAM.
When combined into the Cerebras CS-3 system, the company claims the chip is capable of 125 peak AI petaflops. That can be scaled up to 2,048 CS-3s in a single cluster, for a purported 256 exaflops of AI compute.
"We didn't build it big because big was easy or big was sexy, but not only are we 52 times bigger [than an Nvidia GPU], but we bring 52 times the number of cores, more than 800x on-chip memory, and more than three orders of magnitude more memory bandwidth and communication bandwidth," Cerebras' VP of product Dr. Andy Hock told DCD.
Compared to the WSE-2, "it's the same biggest square we can cut out of the biggest circular primary silicon we can buy," Hock said. "It's going to be 2× the performance for AI work - within the same power envelope and price."
The company has never officially disclosed the price of its chips, but they are believed to cost around $2-3 million. Hock said that a single CS-3 uses 23kW.
Alongside the new chip, the company said that it had completed the construction of Condor Galaxy 2, its second cloud supercomputer built as part of a $900 million deal with G42.
"We're now building the third Condor Galaxy machine," Hock said, with this one featuring CS-3s.
Cerebras has also partnered with Qualcomm to deploy the latter company's Cloud AI 100 Ultra chips for inference work next to the training supercomputers.
Hock said the company was specifically training models to perform well on the inference chips, which would "eliminate a large fraction of unnecessary computation of multiply by zero operations."