[ad_1]
The design can run a large neural network more efficiently than interconnected GPU banks. But chip manufacturing and commissioning is a challenge, requiring new methods to record the characteristics of silicon, a design that includes redundancies to account for manufacturing defects, and a new water system to keep the giant chip cool.
To build a set of WSE-2 chips capable of running record-sized AI models, Cerebras had to solve another engineer’s challenge: how to enter and exit data from the chip efficiently. Conventional chips have their memory on board, but Cerebras developed an off-chip memory box called MemoryX. The company also created software that allows the neural network to be partially stored in that memory outside the chip, changing only the computations of the silicon chip. And he built a hardware and software system called SwarmX, which connects everything together.
“They can improve the scalability of training in addition to what is currently being done, to enormous dimensions,” he says Mike Demler, A senior analyst at The Linley Group and Microprocessor report.
Demler says it’s still unclear how many markets there will be for the cluster, especially since some potential customers are already designing their own specialized chips at home. He adds that in terms of actual chip performance, speed, efficiency and cost, they are still not clear. Cerebras has not yet published benchmark results.
“There’s impressive engineering in new MemoryX and SwarmX technologies,” Demler says. “But just like the processor, they are very specialized things; it makes a lot of sense to train the biggest models. “
Cerebras ’chips have so far been taken up by labs that need supercomputing power. Initial clients include Argonne National Labs, Lawrence Livermore National Lab, pharmaceutical companies such as GlaxoSmithKline and AstraZeneca, and what Feldman describes as a “military intelligence” organization.
This shows that the Cerebras chip can be used more than feeding neural networks; the calculations made by these laboratories are massive parallel mathematical operations. “And there’s always a thirst for more computing power,” Demler says, adding that the chip could become important for the future of supercomputing.
David Kanter, analyst Real World Technologies and executive director MLCommonsThe organization, which measures the performance of different AI algorithms and hardware, says it sees a much larger market for future AI models in general. “In general, I tend to believe in data-based ML, so we want larger data sets that will allow us to build larger models with more parameters,” says Kanter.
[ad_2]
Source link
