Unverified50% confidenceFactExact time
FP8由NVIDIA在Hopper架构（H100）中引入硬件原生支持，随后在Ada Lovelace（RTX 40系）和Blackwell（RTX 50系）架构中持续完善
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
LingBot-Video 1.3B 选择性FP8量化实测：采样提速22%
redditr/comfyui7/11/2026
Related Claims
UnverifiedRTX 5080搭载NVIDIA最新一代Blackwell GPU架构，相较于上一代Ada Lovelace架构在Tensor Core密度和FP8精度计算上有显著提升77% similarVerifiedFP8由NVIDIA在H100架构中首次引入硬件支持，分为E4M3和E5M2两种子类型76% similarUnverifiedRTX 40系列及更新架构的GPU内置了INT8和FP8运算的硬件加速单元75% similarVerifiedFP8是NVIDIA Hopper架构（H100）原生支持的8-bit浮点格式，相比INT8量化保留了更好的动态范围74% similarUnverifiedV100所属的Volta架构不具备原生FP8硬件加速支持，FP8原生支持于Hopper架构的H10073% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/494666API
curl https://kongchang.com/api/v1/knowledge/claims/494666MCP
get_claim(id=494666)