Accelerate AI in the public interest with the most powerful solution for AI computing from Cerebras Sytems

Industry Challenge:

AI has the potential to revolutionize government services, research, policy, and security. But public datasets are massive, and applications require rapid processing to deliver insights in time.

The Cerebras CS-2 delivers more AI compute in a single device than any other, to accelerate large model training and inference at the scale of government applications by orders of magnitude, reducing the cost of research and delivering instant answers in production.

Use Case

Language and time series modeling

Innovation for language and time series AI models are driving important applications for public sector work, ranging from text analysis to translation, to healthcare research and predictive maintenance. But state of the art models in this domain need weeks or months to train even on large GPU clusters.

Researchers need not just more compute, but smarter, purpose-built compute to accelerate training and production inference processing. With the CS-2, researchers can leverage a revolutionary 850,000 core wafer-scale processor to reduce wall-clock computation time by many 10s to 100s of times beyond contemporary GPU systems, all within a single device.

Use Case

Signal and graph processing

Large scale data analytics tasks like radio frequency or other signal processing and graph analysis have compute – memory – communication demands that far exceed the capabilities of traditional, general purpose processors like CPU and GPU.

The CS-2 delivers greater sparse compute, on-chip memory and network bandwidth than any other system — with this, researchers can model and analyze real-world size volumes of signals or graph data orders of magnitude faster than any small, legacy chip solution.

Use Case

Image and video processing

AI has revolutionized how we process image and video data to improve classification and detection for a wide range of environmental, agricultural, and security applications. However, developers are often restricted to use small or low resolution data and constrained by long training and inference times using small chips built for other work.

Because of the WSE’s massive AI compute with fast on-chip memory and interconnect, the CS-2 enables multi-megapixel image processing, large model training in hours rather than days or weeks, and orders of magnitude faster inference to keep the analyst ahead of the data feed rather than reacting days or weeks later.

Integrating Cerebras technology into the Lawrence Livermore National Laboratory supercompute infrastructure enabled us to build a truly unique compute pipeline with massive computation, storage, and thanks to the Wafer Scale Engine, dedicated AI processing.

Bronis de Supinski

CTO, Livermore Computing @ Lawrence Livermore National Laboratory