对 DeepInfer 发布的一条客观批评:异构优化缺位
有人扒 DeepInfer 的短板:不做异构优化,但有 3FS 和 Dynamo 兜底,还说 4 千张 GB200 就能追平 DeepSeek 吞吐。
有人对 DeepSeek 的 DeepInfer 推理方案提出批评,指出它没有针对异构硬件做任何优化。批评者也承认 DeepSeek 开源了 3FS 文件系统,NVIDIA Dynamo 或许能承担负载均衡工作。文中估算:如果有 3-4 千张 GB200,吞吐量可以接近 DeepSeek 官方报告的服务器数据。
One of the few objective criticisms of DeepInfer release In fairness, they do provide 3FS, they don't optimize for heterogenous anything, and perhaps Dynamo would cut it for load balancing. If you have 3-4K GB200s, you can afford to ≈match reported DeepSeek server throughput.