Cerebras details wafer-scale deployment of a trillion-parameter model
What happened
Cerebras laid out how it serves a trillion-parameter model across 59 WSE-3T wafers, using the roughly 3-trillion-parameter open model Kimi K3 as the example. Its 69 linear-attention layers and 24 dense layers are spread over the wafers, reaching 1,000 to over 2,000 tokens per second. Each wafer holds 44 GB of on-chip memory, with one layer taking about 16 GB. Weights stay resident on-chip; only activations cross wafers, over Ethernet-based RDMA at about 2 microseconds per hop.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageCerebras Deploys Trillion-Parameter AI Models
Cerebras 披露在 WSE-3T 晶圆上部署万亿参数模型的方法：以约 3T 参数的开源模型 Kimi K3 为例，将 69 层线性注意力层与 24 层稠密层分布到 59 片晶圆，实现 1000-2000+ tokens/s 的推理速度。每片晶圆有 44 GB 片上内存，单层约 16 GB，模型权重常驻片上、仅激活值经 RDMA over Ethernet 在晶圆间传递，每跳约 2 微秒。
Heat over time
Not enough continuous observations to draw a trend yet.