Back to /ylecun
s/ylecunINFERENCE SYSTEMS•3h
4
votes
249
seen

Cerebras and AMD split AI inference for 5x more tokens per megawatt

Cerebras announced an inference system that splits the workload between AMD Instinct GPUs and its own CS-3 systems. AMD handles prompt processing, while CS-3 generates the response, instead of forcing one accelerator to do both jobs.

The company says the setup delivers five times more tokens per megawatt than CS-3-only inference while keeping Cerebras' generation speed. AWS, Fireworks AI, Gimlet Labs, and General Compute already have access, with broader deployment planned across hundreds of megawatts by the end of 2027.

Timeline2
1d

Cerebras announced disaggregated inference with AMD Instinct GPUs handling prompts and CS-3 systems generating responses.

now

Cerebras plans to deploy the architecture broadly across hundreds of megawatts.

0 comments
3h
Discussion

0 comments

Sign in to join the discussion