Unverified50% confidenceBenchmarkExact time
On a single H100, FuTransformer's prefill path is up to 1.48x faster than compiled PyTorch FP8, reaching 1.29 PFLOP/s.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/19/2026
First Seen
Valid until: 11/2/2026
Sources
Related Entities
Related Claims
UnverifiedH100的Transformer Engine支持FP8精度，张量核心在FP8模式下峰值算力可达3958 TFLOPS，FP16为1979 TFLOPS77% similarVerifiedNVIDIA's H100 features a Transformer Engine specifically designed for AI workloads, delivering nearly 4,000 TFLOPS of compute at FP8 precision, approximately 3 times the improvement over the A100.75% similarUnverifiedFuTransformer's INT4 decode is 31% to 51% faster than its own FP8 path.75% similarUnverifiedThe Hopper architecture of the H100 introduced a dedicated Transformer Engine supporting FP8 precision computation.72% similarUnverifiedAcross 8B and 70B shapes and batch sizes from 1 to 32, FuTransformer's INT4 decode is 1.3 to 2.4x faster than the faster of PyTorch's BF16 and INT4 blocks.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065906API
curl https://kongchang.com/api/v1/knowledge/claims/1065906MCP
get_claim(id=1065906)