Unverified50% confidenceFactExact time
FP8 training schemes typically use E4M3 for the forward pass and E5M2 for the backward pass.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFP8 has two main sub-formats — E4M3 (4-bit exponent + 3-bit mantissa) and E5M2 (5-bit exponent + 2-bit mantissa) — optimized for forward propagation and backward propagation respectively.74% similarVerifiedFP8 was first introduced with hardware support by NVIDIA in the H100 architecture, with two subtypes: E4M3 and E5M2.66% similarUnverifiedFP8 is a low-precision computing format natively supported by the Blackwell architecture, including two variants—E4M3 and E5M2—and the Tensor Core's FP8 compute throughput is twice that of FP16.65% similarUnverifiedIn mixed-precision training, the forward pass uses FP16 to leverage Tensor Core acceleration while parameter updates retain FP32 to maintain numerical stability, which has become standard practice for large model training.63% similarUnverifiedFP8 is a data type natively supported by the NVIDIA Blackwell architecture, available in two formats: E4M3 and E5M2.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/832463API
curl https://kongchang.com/api/v1/knowledge/claims/832463MCP
get_claim(id=832463)