Unverified50% confidenceBenchmarkExact time
FuTransformer's INT4 decode is 31% to 51% faster than its own FP8 path.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/19/2026
First Seen
Valid until: 11/2/2026
Sources
Related Entities
Related Claims
UnverifiedAcross 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.77% similarUnverifiedOn a single H100, FuTransformer's prefill path is up to 1.48x faster than compiled PyTorch FP8, reaching 1.29 PFLOP/s.75% 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 A100.67% similarUnverifiedFuTransformer is a CUDA implementation targeting NVIDIA Hopper GPUs, with a separate FP8 prefill path and INT4 decode path designed for Llama-style Transformer blocks.65% similarVerifiedBlackwell's second-generation Transformer Engine adds FP4 precision support, which in theory can improve Transformer model training throughput by 2–4x compared to the H100.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065911API
curl https://kongchang.com/api/v1/knowledge/claims/1065911MCP
get_claim(id=1065911)