DARPA selects co-packaged optics tech to advance AI computing

03 Apr 2025

Large chips from Cerebras Systems and co-packaged optics from Ranovus to be involved.

Cerebras Systems, a developer of accelerating generative AI processes, has won a contract from the U.S. Defense Advanced Research Projects Agency (DARPA), for the development of a high-performance computing system.

The Cerebras system will combine its wafer scale technology and co-packaged optics developer Ranovus’ wafer scale optical technology to achieve what the partners call “several orders of magnitude better compute performance at a fraction of the power draw.”

“By combining wafer scale technology and co-packaged optics interconnects, 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 Andrew Feldman, co-founder and CEO of Cerebras.

“Building on the successes of the DARPA’s Digital RF Battlespace Emulator (DRBE) program, where Cerebras is currently executing the third phase and delivering a leading-edge RF emulation supercomputer, in this new initiative Cerebras and its partner Ranovus will deliver the industry-first wafer-scale photonic interconnect solution,” added Feldman.

Solving the memory and communication bottlenecks

Two of the fundamental challenges faced by today’s computer systems are memory and communication bottlenecks, say the partners: “The compute requirement is growing at a much faster rate than either memory or IO technologies can keep up with.” Cerebras adds that its work in wafer scale integration “has already closed the memory bandwidth gap.” The firm’s Wafer-Scale Engine has 7,000 times more memory bandwidth than typical GPUs, and as a result can deliver the world’s fastest AI inference and the world’s fastest molecular simulations ever conducted, the company claims.

Cerebras adds that, As part of the new DARPA effort, its approach “will solve the communication bottleneck by integrating advanced co-packaged optics interconnects from Ranovus, enabling compute performance impossible to achieve today even in the largest of supercompute clusters.”

The solution is expected to deliver compute at a fraction of the power consumed by GPUs linked together with traditional switches. Today’s switches and their optical interconnects are amongst the most power-hungry components in a large AI or simulation cluster. Integrating the optics into the wafer scale package provides unmatched power efficiency, says Cerebras.

The impact of this technology extends beyond national defense, serving dual-use applications in both Department of Defense (DoD) and commercial sectors. In particular, it holds immense potential for doing real-time AI directly from information processing sensors, as well as real-time simulations of battlefields, and military and commercial robotics.

Hamid Arabzadeh, CEO and founder of Ranovus, said, “Our wafer-scale co-packaged optics platform will deliver 100 times the capacity of current co-packaged optics solutions while significantly enhancing the energy efficiency of AI clusters. This partnership will establish a new industry standard for supercomputing and AI infrastructure, addressing the rising demand for data transmission and processing and paving the way for the next generation of military and commercial simulations and applications.”