谷歌TPU跑Kimi比英伟达GPU快57%！用的还是DeepSeek推理框架

AI Digest

谷歌TPU v7跑Kimi K3比英伟达GPU快57%，得益于DeepSeek框架与Inferact的megakernel技术优化。

Google's TPU v7 outperforms NVIDIA GPU by 57% on Kimi K3 using DeepSeek's DSpark framework and Inferact's megakernel optimization.

Key points

• TPU v7推理速度比GB200快57%，带宽利用率接近硬件峰值 TPU v7 achieves 57% higher throughput than NVIDIA GPU on Kimi K3
• megakernel技术将多个小程序合并为大程序，消除内存带宽浪费 Megakernel merges multiple kernels into one program to eliminate bandwidth waste
• DSpark框架通过推测解码提升推理效率，单步耗时约8.5毫秒 DSpark framework boosts inference efficiency via speculative decoding
• TPU硬件特性支持跨层权重预取，实现计算与数据搬运并行 TPU's architecture enables cross-layer weight prefetching
• vLLM原班人马创业公司Inferact开源了TPU优化方案 Inferact, a vLLM team spin-off, open-sourced TPU optimizations

Takeaway: 软件优化可突破硬件性能瓶颈，TPU速度优势源于深度定制的推理框架。 / Software optimization can unlock hardware performance limits, with TPU's speed advantage stemming from tailored inference frameworks.

Why it matters 展示软件创新对硬件性能的颠覆性影响，具有实际工程参考价值。