DeepSeek 硬件利用率质疑
精选理由
DeepSeek 的硬件利用率数据引发争议，可能影响其技术可信度。
DeepSeek 被质疑宣称达到 GEMM 硬件利用率 99.8% 和 MegaMoE 98% 后直接开源内核。有观点认为其小型单 GPU 模型可能是基于华为 Ascend 平台开发。这一质疑涉及 DeepSeek 的技术透明度和硬件适配能力。
原文 · Teortaxes
Btw I don't think that DeepSeek «achieved 99.8% of the hardware limit on GEMM and 98% on MegaMoE» and then just open sourced kernels before training anything usable end to end. It wouldn't surprise me if at least the rumored small single-GPU model is Ascend-born.
