PCIe显卡被低估了!内核补齐+通信重构,DeepSeek推理吞吐翻近7倍
中文摘要
通过内核补齐与通信重构,DeepSeek在PCIe显卡上的推理吞吐量提升近7倍,1.5台6000D性能超越1台B300。
English Summary
Kernel and communication optimizations boost DeepSeek inference throughput on PCIe GPUs nearly 7x, allowing 1.5 6000Ds to outperform a B300.
Original Excerpt
1.5台6000D跑赢1台B300!