Cerebras Launches Disaggregated Inference for 5x Throughput per Megawatt

·2026.11.18 09:56

Key point

The production system pairs GPUs for prefill with Cerebras WSEs for decode to deliver five times more tokens per megawatt without sacrificing generation speed.

1 / 2

Details

  • Cerebras has deployed disaggregated inference in production, using GPUs for prompt processing (prefill) and Cerebras Wafer Scale Engines for response generation (decode).
  • Measured production results show a 5x increase in throughput per megawatt compared to a CS-3-only aggregated configuration, with no loss in token generation speed.
  • The architecture leverages Cerebras' high on-chip SRAM capacity and memory bandwidth, which are 350x and 5,400x higher than a typical GB300 GPU, to optimize the memory-bandwidth-bound decode stage.
  • Early-access partners including Amazon Web Services, Fireworks AI, Gimlet Labs, and General Compute already have access to the solution.
  • Cerebras plans to deploy the system across hundreds of megawatts in its Inference Cloud and partner datacenters by the end of 2027.
  • Infrastructure partners can adopt the solution via an end-to-end software stack or a decode-only SDK, using open-source engines like vLLM or SGLang for the prefill stage.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.