Verified65% confidenceBenchmarkExact time
NVIDIA 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
3
Sources
65%
Confidence
Long-term
Relevance
7/23/2026
First Seen
Sources
Related Claims
UnverifiedNVIDIA 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82% 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82% similarUnverifiedNVIDIA A100、Google TPU v4等专用AI训练芯片针对Transformer的矩阵运算模式进行了专项优化81% similarUnverifiedNVIDIA随H100推出Transformer Engine，为每个张量维护历史激活值最大绝对值统计并动态调整缩放因子80% similarUnverifiedH100的Hopper架构引入了专门的Transformer Engine，支持FP8精度计算79% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/599795API
curl https://kongchang.com/api/v1/knowledge/claims/599795MCP
get_claim(id=599795)