Unverified50% confidenceFactExact time
现代GPU如NVIDIA H100、H800原生支持FP8计算，理论上可实现约2倍吞吐量提升
1
Sources
50%
Confidence
Long-term
Relevance
7/10/2026
First Seen
Sources
腾讯混元3开源：295B MoE架构全解析，256K上下文成最大短板
bilibili鲲鹏Talk7/6/2026
Related Claims
UnverifiedNVIDIA H100 GPU原生支持FP8格式，其吞吐量相较FP16提升约2倍79% similarUnverified英伟达H100、H800在FP8精度下的理论推理吞吐相比A100提升了3-4倍73% 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73% similarUnverifiedFP32到FP16的转换在现代GPU的Tensor Core上可实现2-4倍的吞吐量提升71% similarUnverified英伟达GB300 Blackwell Ultra GPU集成两颗GPU芯片与一颗Grace CPU，通过NVLink-C2C互联，内存带宽可达8TB/s以上，推理性能相比H100提升4至5倍70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/465619API
curl https://kongchang.com/api/v1/knowledge/claims/465619MCP
get_claim(id=465619)