How Cerebras deploys trillion-parameter models
Cerebras describes mapping a roughly 3-trillion-parameter Kimi K3 model across 59 WSE-3 wafers: one dense-attention layer per wafer to reserve on-chip memory for the KV cache, and two linear-attention layers per wafer. The pipeline passes activations between wafers over RDMA Ethernet; the post estimates about 2 microseconds per hop.
These are vendor design details and illustrative capacity calculations, not an independent performance evaluation. The article’s 1,000–2,000+ tokens/second figure is a stated target; the HN discussion permalink currently resolves to the story listing without readable comments, so I’m not attributing reader reactions.