Unverified50% confidenceFactExact time
H100的Hopper架构引入了专门的Transformer Engine，支持FP8精度计算
1
Sources
50%
Confidence
Long-term
Relevance
7/20/2026
First Seen
Sources
Related Claims
UnverifiedH100 GPU基于Hopper架构，其Transformer引擎能在FP8和FP16精度间动态切换，将吞吐量提升至A100的约2-3倍84% similarUnverifiedFP8精度（Hopper架构H100引入）可将Transformer训练吞吐提升近2倍，若训练超大模型且框架支持Transformer Engine，H100相比A100的实际训练加速远超纸面算力差距83% similarUnverifiedH100的Transformer Engine支持FP8精度，张量核心在FP8模式下峰值算力可达3958 TFLOPS，FP16为1979 TFLOPS82% similarVerifiedNVIDIA H100采用专为AI工作负载设计的Transformer Engine，在FP8精度下实现近4000 TFLOPS算力，比A100提升约3倍79% similarVerifiedNVIDIA Hopper架构（H100/H200）首次加入FP8 Tensor Core，并引入Transformer Engine自动管理精度转换77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/573242API
curl https://kongchang.com/api/v1/knowledge/claims/573242MCP
get_claim(id=573242)