A New Chip Cluster Will Make Many AI Models Possible
The scheme can run a large neural network more efficiently than GPU banks combined. But making and running the chip is a challenge, requiring new methods for etching silicon shapes, a design with redundancies to account for manufacturing errors, and a novel system. of water to prevent the giant chip.
In order to build a cluster of WSE-2 chips capable of running record-sized AI models, Cerebras must solve another engineering challenge: how to capture data in and out of the chip. . Regular chips have their own on-board memory, but Cerebras makes an off-chip memory box called MemoryX. The company also developed software that allows a neural network to be stored in that memory without a chip, with only the computations closed on the silicon chip. And it built a hardware and software system called SwarmX that put it all together.
“They can improve the level of training skills on many dimensions beyond what anyone is doing today,” he said. Mike Demler, a senior analyst at The Linley Group and a senior editor at The Microprocessor Report.
Demler said it’s not yet clear how much of a market there will be for the cluster, especially since some potential customers are already designing their own, more specialized in-house chips. He added that the actual making of the chip, in terms of speed, efficiency, and cost, is not yet clear. Cerebras has not published any benchmark results to date.
“There’s a lot of amazing engineering to the new technology in MemoryX and SwarmX,” Demler said. “But like the processor, it’s a very specialized thing; it makes sense for training on most models.”
Cerebras chips to date have been used in labs that require supercomputing power. Early customers included Argonne National Labs, Lawrence Livermore National Lab, pharma companies including GlaxoSmithKline and AstraZeneca, and what Feldman described as “military intelligence” organizations.
This shows that the Cerebras chip can be used more to run neural networks; calculations run in labs involving the same widespread parallel mathematical operation. “And they’re always thirsty for more computing power,” said Demler, who added that the chip could be crucial for the future of supercomputing.
David Kanter, a researcher at Real World Technologies and executive director of MLCommon, an organization that measures the development of a variety of AI algorithms and hardware, said it sees a future market for more AI models in general. “I usually believe in data -centric ML, so we want more datasets to be able to build more models with more parameters,” Kanter said.