Cerebras在WSE-3T芯片上部署万亿参数AI模型

Cerebras发布技术文章，介绍了在其WSE-3T晶圆芯片上部署万亿级参数AI模型的方案1。该公司以开源模型Kimi K3为例进行了详细说明，展示了大规模模型在专用硬件上的实现路径1。

Kimi K3模型拥有约3万亿参数，包含69个线性注意力层和24个密集层1。单块WSE-3T晶圆拥有44GB片上内存，部署该模型需要总共59块晶圆（其中35块用于线性层、24块用于密集层），每层模型在采用4比特权重量化的情况下需要约16GB内存1。芯片间通过RDMA over Ethernet传输激活值，延迟约为2微秒每跳1。通过跨多块芯片的管道架构分配模型权重，该系统可实现每秒1000-2000+token的推理速度1。

在性能评测中，Kimi K3在Terminal-Bench 2.1上的得分为88.3%（相比之下GPT-5.6 Sol为88.8%），在BrowseComp上的得分为91.2%（GPT-5.6 Sol为90.4%）1。Cerebras采用管道并行架构来优化系统吞吐量，通过保持管道满载不同请求的方式进一步提升性能1。

Cerebras has published technical details on deploying trillion-parameter artificial intelligence models across its WSE-3T wafer-scale processors, demonstrating a practical approach to handling extremely large language models through distributed inference architecture 1. The company used the open-source Kimi K3 model, which contains approximately three trillion parameters including 69 linear attention layers and 24 dense layers, as a case study for its deployment strategy 1.

The technical solution leverages pipelined parallelism to distribute model weights across multiple chips, with each WSE-3T wafer featuring 44 gigabytes of on-chip memory 1. Deploying the full Kimi K3 model required a total of 59 wafers—35 dedicated to linear layers and 24 for dense layers—with individual layers consuming approximately 16 gigabytes of memory when using 4-bit weight quantization 1. Communication between chips occurs through RDMA over Ethernet with a latency of approximately 2 microseconds per hop 1.

The system achieves inference speeds of 1,000 to 2,000 tokens per second through continuous pipeline optimization that maintains full utilization by processing multiple concurrent requests 1. Performance benchmarks show Kimi K3 scoring 88.3 percent on Terminal-Bench 2.1 (compared to 88.8 percent for GPT-5.6 Sol) and 91.2 percent on BrowseComp (compared to 90.4 percent for GPT-5.6 Sol), demonstrating competitive capability despite the model's scale 1.