Unverified50% confidenceBenchmarkExact time
On the H100, FA4's forward pass can achieve an MFU of over 75% of the theoretical peak.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/1/2026
First Seen
Valid until: 12/30/2026
Sources
Related Entities
Related Claims
UnverifiedThe H100's fourth-generation Tensor Core natively supports FP8, with a peak compute of 3958 TFLOPS, which is twice that of FP16.68% similarUnverifiedThe theoretical inference throughput of the NVIDIA H100 and H800 at FP8 precision is 3–4 times higher than that of the A100.66% similarUnverifiedBlackwell 的第五代 NVLink、HBM3e 显存和专用 FP4 张量核心使其在推理密集型场景下相较 H100 可获得2-4倍吞吐量提升64% similarVerifiedThe H100 GPU adopts NVIDIA's Hopper architecture, with a single-card peak compute of 3,958 TFLOPS (FP16 precision).62% similarUnverifiedRunning the model in full precision requires several H100s.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/966304API
curl https://kongchang.com/api/v1/knowledge/claims/966304MCP
get_claim(id=966304)