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 expertise 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 intention of the platform is to deal with two main challenges: reminiscence and communication bottlenecks. Compute requirement are rising quicker than both reminiscence or IO applied sciences can sustain with. Cerebras’ work in wafer scale integration, delivered through dinner plate-sized processor, is designed to handle the reminiscence bandwidth hole. The Cerebras Wafer-Scale Engine has 7,000 instances 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 expertise and digital and silicon photonics built-in circuits — have set new benchmarks for optical interconnects, in line with Ranovus.
The affect of this expertise extends past nationwide protection, serving dual-use functions in each Division of Protection and business sectors. Particularly, it holds immense potential for doing real-time AI immediately from info processing sensors, in addition to real-time simulations of battlefields, and army and business robotics, the businesses mentioned.
“By combining wafer scale expertise and co-packaged optics interconnects, Cerebras will ship a platform able to real-time, high-fidelity simulations for probably the most difficult bodily atmosphere simulations and the biggest 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 at present executing the third part and delivering a modern RF emulation supercomputer, on this new initiative Cerebras and its companion Ranovus will ship the industry-first wafer-scale photonic interconnect resolution.”
As a part of this DARPA effort, Cerebras mentioned it’s going to handle the communication bottleneck by integrating superior co-packaged optics interconnects from Ranovus, enabling compute efficiency “unimaginable to realize right this moment even within the largest of supercompute clusters. Not solely will this resolution be orders of magnitude quicker than right this moment’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. At the moment’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 offers unmatched energy effectivity.”
“By fixing these basic issues of compute bandwidth, communication IO and energy per unit compute by means of Cerebras’ wafer scale expertise plus optical integration with Ranovus co-packaged optics, we are going to unlock options to among the most complicated issues within the realm of real-time AI and bodily simulations — options which can be right this moment completely unattainable,” mentioned Feldman. “Staying forward of our rivals with superior, energy environment friendly compute that permits quicker AI and quicker simulations is vital for US protection, in addition to for the home business 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 instances 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 way in which for the subsequent technology of army and business simulations and functions.”
