Cerebras' Wafer-Scale Engine 3, the "Fastest AI Chip in the World," Aims at Rapid Gen AI Training
With four billion transistors, the company's latest wafer-scale chip is a beast β twice as fast as its predecessor at the same power draw.
Cerebras, which lays claim to having developed the fastest single chip for artificial intelligence workloads, has announced its latest-generation Wafer-Scale Engine 3 β delivering, it says, twice the performance of its predecessor at the same power draw and price.
"When we started on this journey eight years ago, everyone said wafer-scale processors were a pipe dream. We could not be more proud to be introducing the third-generation of our groundbreaking wafer-scale AI chip," Andrew Feldman, chief executive officer and co-founder of Cerebras, says of the company's latest launch.
"WSE-3 is the fastest AI chip in the world," Feldman claims, "purpose-built for the latest cutting-edge AI work, from mixture of experts to 24 trillion parameter models. We are thrilled to bring WSE-3 and CS-3 to market to help solve todayβs biggest AI challenges."
Cerebras' breaking of performance records comes from the unusual design of its chips. Where most companies build multiple small chips per silicon wafer, its Wafer-Scale Engine is exactly what it sounds like: a single giant chip, built one per silicon wafer. As a result, it can do in one chip what it takes its rivals multiple chips to achieve β and the WSE-3 is the company's most ambitious design yet.
The WSE-3 is made up of four trillion transistors across 900,000 cores, delivering a claimed 125 petaflops of performance, and includes 44GB of on-chip static RAM (SRAM). The chip powers the company's latest CS-3 "AI supercomputer," which supports up to 1.2PB of external memory and can be clustered with up to 2,048 nodes β a configuration that is capable of training the 70-billion-parameter open-source Llama model in a single day, the company claims.
Cerebras has already announced a major win for the CS-3, with a partnership that will see Abu Dhabi-based G42 building its Condor Galaxy 3 cluster using 64 CS-3s β delivering eight exaflops of compute using nearly 58 million cores.
Pricing for the WSE-3 and CS-3 is available on application, with more information to be found on the Cerebras website.
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.