Cerebras Puts Disaggregated Inference Into Production With GPUs

Unite.AI
Read full story· 4 min
Cerebras Systems deployed a disaggregated inference setup using AMD Instinct GPUs for prompt processing and CS-3 wafer-scale systems for response generation, achieving five times the throughput per megawatt compared to CS-3 alone without slowing token generation. This production deployment enhances Cerebras Inference Cloud's capacity for fast interactive AI workloads by optimizing resource use across hardware. The approach separates prefill and decode stages to leverage the strengths of GPUs and

What happened

On November 17, 2026, Cerebras Systems announced that it had deployed a production disaggregated inference setup in its Inference Cloud. In this setup, AMD Instinct GPUs handle prompt processing, while Cerebras CS-3 wafer-scale systems generate responses. The company measured a fivefold increase in throughput per megawatt compared to using CS-3 alone, with no reduction in token generation speed. The deployment follows earlier tests and now includes a reference architecture and software options for infrastructure partners.

Key details

  • The production measurement was taken on November 14, 2026, using a shared GPT-OSS-120B endpoint.
  • The reference architecture allows partners to co-locate their own prefill accelerators with Cerebras systems on a unified network.
  • Partners can choose between a full software stack or a decode-only SDK for integration.

Why it matters

Operators of AI workloads, including cloud providers and infrastructure partners, can now serve more interactive AI requests within the same power budget, increasing capacity and potential revenue. The deployment allows partners to integrate Cerebras' high-throughput decode systems with their own prompt-processing hardware, offering flexibility and faster time to market. Customers using Cerebras Inference Cloud or partner services may see faster, more scalable AI response times for interactive applications without increased energy use.


More in Chips & Compute