Tsar Chips with Integrated Photonics: Cerebras Systems and Ranovus Selected by DARPA to Build Next-Gen Computing Platform

AI startup Cerebras Systems has been selected by the US Defense Advanced Research Projects Agency (DARPA) to develop a next-generation high-performance computing system. Cerebras will combine its own AI accelerators with Ranovus photonic CPO interconnects to provide high performance with low power consumption, Cerebras’ press center reports.

The combination of the two companies’ technologies will enable real-time simulation of complex physics processes and execution of large-scale AI tasks. Given the success of the DARPA Digital RF Battlespace Emulator (DRBE) program, in which Cerebras is already developing an advanced RF emulation supercomputer, Cerebras and Ranovus were selected for the new initiative, which will combine Cerebras’ computing products with Ranovus’ industry-first photonic interconnects.

The solution is extremely relevant, since two key issues for modern computing systems are problems with memory and data exchange between accelerators and other server infrastructure – computing needs are growing faster than the capabilities of memory or input-output IO systems. According to Cerebras, its WSE chips have 7 thousand times greater throughput than classic accelerators, which provides the world’s fastest inference and the fastest modeling of molecular processes.

Image source: Cerebras

As part of DARPA’s new plan, startup Cerebras will use the Ranovus interconnect, which will enable it to achieve performance that even the largest supercomputer clusters of today cannot achieve. At the same time, the energy consumption will be significantly lower than that of the most advanced solutions using switches, which are among the most energy-intensive components in modern AI systems or supercomputers.

It is claimed that the combination of new technologies from the two companies will enable solutions to the most complex problems in real time, be it AI or complex modeling of physical processes, at a level that is unattainable today. It is emphasized that staying ahead of the competition is an urgent need for the US defense, as well as the local commercial sector. In particular, this opens up huge opportunities for the operation of AI in real time – from processing data from sensors to simulating combat operations and controlling combat or commercial robots.

Ranovus said the Wafer-Scale Co-Packaged Optics platform is 100 times more productive than similar current solutions, which allows for a significant increase in the efficiency of AI clusters, and is significantly more energy efficient than competitors’ products. The partnership between the companies will set a new standard for supercomputer and AI infrastructure, addressing the growing demand for data transmission and processing and enabling the implementation of next-generation military and commercial modeling.

In addition to being used for the US military, Cerebras’ giant AI chips are also used by defense agencies in other countries. In the spring of 2024, for example, it was reported that the company’s products would help train AI for the German military.