SiliconCerebras · AMD
Cerebras put disaggregated inference into production, pairing AMD GPUs with its wafer chips
5xthroughput per megawatt against CS-3 alone, in Cerebras's production measurement
Cerebras said on 17 November that disaggregated inference is now in production. AMD Instinct GPUs handle prefill, the reading of the prompt. Cerebras CS-3 systems handle decode, the writing of the answer. Cerebras measured 5x the throughput per megawatt of CS-3 alone. It says output speed did not drop. Partners can use its full stack or a decode-only SDK. Cerebras plans to deploy it across hundreds of megawatts by the end of 2027.
Cerebrasfrom the source itself