Cerebras Puts Disaggregated Inference Into Production With GPUs
Reported by Cerebras, Unite.AI, Hacker News and 1 more outlets4 sources ↓
Cerebras says its production inference cloud now uses GPUs for prompt processing and Cerebras systems for response generation. In a measured deployment, the setup delivered five times the throughput per megawatt of a CS-3-only configuration without reducing token generation speed.
Key points
- AMD Instinct GPUs process prompts while Cerebras CS-3 systems generate responses
- Cerebras reports five times the throughput per megawatt of a CS-3-only setup
- The measurement used a shared GPT-OSS-120B endpoint on November 14
- AWS, Fireworks AI, Gimlet Labs and General Compute have early access
- Cerebras plans deployments across hundreds of megawatts by the end of 2027
Sources
Details
Cerebras says the production deployment is running on its Inference Cloud, with AMD Instinct GPUs handling prompt processing and CS-3 systems generating responses. The company says moving prompt processing to GPUs frees its systems to handle response generation.
The reported fivefold throughput-per-megawatt result came from a shared GPT-OSS-120B endpoint measured on November 14 and was normalized against an aggregated inference baseline. Cerebras says token generation speed did not decline; the articles provide no independent verification of the measurement.
Infrastructure partners can use a Cerebras reference architecture that connects their chosen prompt-processing accelerators with Cerebras systems. The company says partners can integrate either its end-to-end software stack or a decode-only software development kit; AWS, Fireworks AI, Gimlet Labs and General Compute have early access.
Cerebras says it plans to deploy the approach across hundreds of megawatts in its cloud and partner data centers by the end of 2027. This is a stated plan, not a report of capacity already in operation.
Written from coverage as of
Images
- ecosistemastartup
- Hacker News
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh