Cerebras describes distributing model layers across wafers with 44GB of on-chip memory to serve very large models. Its illustrative mapping for the roughly three-trillion-parameter Kimi K3 assumes four-bit weights and an eight-bit KV cache, allocating 35 wafers to linear-attention layers and 24 to full-attention layers, for 59 in total. Weights remain resident while activations travel over Ethernet, and multiple requests fill the pipeline. Cache capacity and request length constrain concurrency. This inference design differs from the external-memory weight-streaming training architecture introduced in 2021, and the stated speeds are not independent benchmark results.
Back to Tech
Cerebras explains multi-wafer inference for trillion-parameter models
Weights stay on-chip while activations move between wafers; the 59-wafer mapping is an illustration with specific quantization assumptions.