AI compute firm Cerebras Methods mentioned it has been awarded a brand new contract from the Protection Superior Analysis Initiatives Company (DARPA) to develop a system combining their wafer scale know-how with wafer scale co-packaged optics of Ottawa-based Ranovus to ship “a number of orders of magnitude higher compute efficiency at a fraction of the ability draw.”
Cerebras mentioned the goal of the platform is to sort out two main challenges: reminiscence and communication bottlenecks. Compute requirement are rising sooner than both reminiscence or IO applied sciences can sustain with. Cerebras’ work in wafer scale integration, delivered by way of dinner plate-sized processor, is designed to handle the reminiscence bandwidth hole. The Cerebras Wafer-Scale Engine has 7,000 occasions extra reminiscence bandwidth than GPUs, and delivers the quickest AI inference and the quickest molecular simulations, in line with the corporate.
Ranovus, in the meantime, focuses on superior co-packaged pptics interconnect for AI/ML workloads in information facilities and communication networks with its Odin platform. The corporate’s IP cores — together with multi-wavelength quantum dot laser know-how and digital and silicon photonics built-in circuits — have set new benchmarks for optical interconnects, in line with Ranovus.
The impression of this know-how extends past nationwide protection, serving dual-use functions in each Division of Protection and industrial sectors. Particularly, it holds immense potential for doing real-time AI instantly from data processing sensors, in addition to real-time simulations of battlefields, and navy and industrial robotics, the businesses mentioned.
“By combining wafer scale know-how and co-packaged optics interconnects, Cerebras will ship a platform able to real-time, high-fidelity simulations for probably the most difficult bodily setting simulations and the most important scale AI workloads, pushing the boundaries of what’s doable in AI and in excessive efficiency computing and AI” mentioned Andrew Feldman, co-founder and CEO of Cerebras. “Constructing on the successes of the DARPA’s Digital RF Battlespace Emulator (DRBE) program, the place Cerebras is presently executing the third section and delivering a modern RF emulation supercomputer, on this new initiative Cerebras and its associate Ranovus will ship the industry-first wafer-scale photonic interconnect resolution.”
As a part of this DARPA effort, Cerebras mentioned it can handle the communication bottleneck by integrating superior co-packaged optics interconnects from Ranovus, enabling compute efficiency “unattainable to attain at this time even within the largest of supercompute clusters. Not solely will this resolution be orders of magnitude sooner than at this time’s state-of-the-art, however it additionally will ship compute at a fraction of the ability consumed by GPUs tied along with conventional switches. Immediately’s switches and their optical interconnects are amongst probably the most power-hungry parts in a big AI or simulation cluster. Integrating the optics into the wafer scale bundle gives unmatched energy effectivity.”
“By fixing these elementary issues of compute bandwidth, communication IO and energy per unit compute by means of Cerebras’ wafer scale know-how plus optical integration with Ranovus co-packaged optics, we’ll unlock options to a number of the most advanced issues within the realm of real-time AI and bodily simulations — options which are at this time totally unattainable,” mentioned Feldman. “Staying forward of our rivals with superior, energy environment friendly compute that permits sooner AI and sooner simulations is important for US protection, in addition to for the home industrial sector.”
“We’re thrilled to collaborate with Cerebras on this groundbreaking innovation,” mentioned Hamid Arabzadeh, CEO and founding father of Ranovus. “Our Wafer-Scale Co-Packaged Optics platform will ship 100 occasions the capability of present Co-Packaged Optics options whereas considerably enhancing the power effectivity of AI clusters. This partnership will set up a brand new {industry} normal for supercomputing and AI infrastructure, addressing the rising demand for information transmission and processing and paving the best way for the following era of navy and industrial simulations and functions.”
