Unverified50% confidenceBenchmarkExact time
B200的FP4算力峰值可达约9PFLOPS（稀疏模式），相比FP16提升可达8倍
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/7/2026
First Seen
Valid until: 10/5/2026
Sources
NVFP4量化实战：NVIDIA模型优化器压缩大模型指南
rss6/26/2026
Related Claims
UnverifiedFP8数据量比FP16压缩一半，H100的FP8 Tensor Core算力达到FP16的两倍（约1979 TFLOPS）73% similarUnverifiedNVIDIA A100的理论FP16峰值算力约为312 TFLOPS，但朴素LLM推理场景下GPU利用率常不足30%73% similarUnverifiedNVIDIA B200于2024年发布，FP16算力达到4.5 PFLOPS级别71% similarUnverifiedA100拥有40/80GB显存和19.5 TFLOPS的FP32性能71% similarUnverified官方提供FP8量化检查点（Hy3-FP8），FP8量化相比FP16可将显存占用和显存带宽需求减半70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/187839API
curl https://kongchang.com/api/v1/knowledge/claims/187839MCP
get_claim(id=187839)