AI BriefingCerebras Explains How to Deploy 3T Parameter Models on WSE-3T Wafers
·2026.10.23 09:00
Key point
The deployment uses 4-bit quantization and pipeline parallelism to keep model weights static on-chip while streaming activations between wafers.
Details
- Cerebras deploys trillion-parameter models like Kimi K3 (~3T parameters) across multiple WSE-3T wafers to achieve 1000-2000+ tokens per second.
- Kimi K3 requires 59 wafers total: 35 for 69 linear-attention layers (two per wafer) and 24 for dense layers (one per wafer), using 4-bit weight quantization and 8-bit KV cache.
- Model weights remain static on individual wafers because each 44 GB wafer holds full 16 GB layers; only activations move between wafers via RDMA over Ethernet at roughly 2 microseconds per hop.
- The system serves concurrent requests by keeping the pipeline full with different requests at each stage, using micro-batches to balance latency and throughput.
- Concurrent request capacity is limited by the remaining on-chip SRAM on dense-layer wafers after accounting for weights, which stores the KV cache for active requests.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
An Interview with OpenAI's Richard Ho on Why Jalapeño Is a Groundbreaking AI Chip