Cerebras wins DARPA contract to build AI system with co-packaged optics – News

Cerebras Systems Inc., the creator of a wafer-scale chip with nearly 1 million cores, has won a hardware development contract from the U.S. Defense Advanced Research Projects Agency.

Reuters reported that the agreement is worth $45 million. As part of the project, Cerebras will develop a new “high-performance computing system” for DARPA. The system will combine the company’s chip with co-packaged optics technology from Ranovus Inc., a Canadian semiconductor startup.

Sunnyvale, California-based Cerebras develops a chip known as the WSE-3 (pictured) that is optimized for artificial intelligence workloads. It’s the size of a silicon wafer and features 4 trillion transistors organized into 900,000 cores. There’s also a 44-gigabyte pool of onboard SRAM memory.

Cerebras positions the WSE-3 as a more efficient alternative to traditional graphics cards. In many cases, the chip’s large onboard SRAM pool allows it to store all the data that AI models generate during processing. This removes the need to move data off the WSE-3 to external RAM. Because onboard SRAM is considerably more power-efficient than external RAM, the WSE-3 can move data to and from its cores using less electricity than graphics cards that have less onboard memory.

The computing system that Cerebras is building for DARPA will also feature “wafer scale co-packaged optics” from Ranovus. According to Reuters, the latter technology will be used to link together multiple WSE-3 chips to create AI clusters.

Data center operators often connect the servers in their AI clusters to one another with fiber-optic cables. Before data can travel over a fiber-optic cable, it has to be turned into light beams. After the light arrives at its destination, it has to be turned back into electrical signals to facilitate processing. Those tasks were historically performed by network devices called pluggable transceivers.

Co-packaged optics technology seeks to remove the need for standalone pluggable transceivers. It does so by integrating a transceiver directly into the processors that power an AI cluster. According to Cerebras, integrating Ranovus’ co-packaged optics with its chips will make customers’ AI clusters more efficient and boost their network capacity in the process. 

“Cerebras will deliver a platform capable of real-time, high-fidelity simulations for the most challenging physical environment simulations and the largest scale AI workloads,” said Cerebras Chief Executive Officer Andrew Feldman.

The new hardware development contract is not the only project that Cerebras is carrying out for DARPA. The chipmaker is also involved in a second research initiative, the DRBE program, that seeks to create a virtual environment for testing wireless communications systems. The project will see Cerebras and its partners build a radio frequency “emulation supercomputer” for DARPA.

In parallel, the chipmaker is building six new data centers equipped with its WSE-3 processors. The facilities are set to include thousands of Cerebras’ CS-3 compute appliances, which each combine one WSE-3 with cooling equipment and other auxiliary components. 

Photo: Cerebras Systems