Cerebras launches disaggregated inference: GPUs prefill, WSE decodes, 5x tokens/MW

AccBalanced · x · 2026-11-18

Cerebras announced disaggregated inference: pairing its Wafer Scale Engines with GPUs, where GPUs handle prompt processing (prefill) and Cerebras systems handle output generation (decode), claiming five times more tokens per megawatt at Cerebras-level speeds. A quoted comment frames this against NVIDIA's reported $20B Groq acquisition (unverified) and notes disaggregated clusters have yet to go live at scale. The infographic explains the decode bottleneck: GPUs stream weights from off-chip HBM per token, while WSE keeps weights in on-chip SRAM, making decode roughly 10x faster since speed is set by how fast weights reach compute.

Related event: Cerebras Launches Disaggregated Inference Pairing GPUs with Waferscale Chips(2 posts)→

Original post →

More from Infra

Infra channel →