Unverified83% confidenceFactTime unknown
FP8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
1
Sources
83%
Confidence
Medium-term (~90 days)
Relevance
-
First Seen
Sources
Related Entities
Related Claims
UnverifiedH100的Hopper架构引入了专门的Transformer Engine，支持FP8精度计算83% similarUnverifiedNVIDIA H100 的 Transformer Engine 将 FP8 训练和推理的吞吐量相比上一代提升近 4 倍81% similarUnverifiedH100 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77% similarUnverifiedBlackwell的第二代Transformer Engine新增FP4精度支持，理论上可将Transformer模型的训练吞吐量相比H100提升2-4倍75% 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74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/468479API
curl https://kongchang.com/api/v1/knowledge/claims/468479MCP
get_claim(id=468479)