Expired50% confidenceSolutionExact time
For NVFP4, the industry generally adopts an asymmetric strategy of forward FP4 and backward BF16, which is also the recommended usage of NVIDIA Transformer Engine.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/14/2026
First Seen
Valid until: 10/12/2026(expired)
Sources
NVFP4强化学习训练：4位量化稳定性挑战与工程实践
Related Claims
UnverifiedThe Transformer Engine of the Hopper architecture can dynamically switch precision between FP8 and FP16 on a per-layer basis.70% similarUnverifiedNVIDIA launched the Transformer Engine alongside the H100, which maintains historical maximum absolute value statistics of activations for each tensor and dynamically adjusts scaling factors.67% similarUnverifiedNVFP4是NF4（Normal Float 4）等学术方案的工程化演进版本66% similarUnverifiedNVIDIA H100 的 Transformer Engine 将 FP8 训练和推理的吞吐量相比上一代提升近 4 倍65% similarPartially VerifiedNVFP4 is a 4-bit floating-point (FP4) quantization format introduced by NVIDIA, designed specifically for the hardware acceleration capabilities of Blackwell architecture GPUs64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/510046API
curl https://kongchang.com/api/v1/knowledge/claims/510046MCP
get_claim(id=510046)