Verified65% confidenceFactExact time
FP8 has two variants: E4M3 (4-bit exponent + 3-bit mantissa), which has higher precision and is suited for forward inference, and E5M2 (5-bit exponent + 2-bit mantissa), which has a larger dynamic range and is suited for gradient computation.
3
Sources
65%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
LingBot-Video 1.3B Selective FP8 Quantization Benchmark: 22% Faster Sampling
redditr/comfyui7/11/2026
Related Claims
UnverifiedFP8 (E4M3/E5M2) occupies 1 byte per parameter and is the mainstream low-precision inference format on current Hopper and Blackwell.68% similarVerifiedThe core principle of quantization is to convert model parameters from high-precision floating-point representations (such as FP32's 32-bit floating point) to low-precision representations (such as INT8's 8-bit integer or INT4's 4-bit integer).67% similarVerifiedNumerical gradient checking approximates gradients using the central difference formula (f(θ+ε) - f(θ-ε)) / (2ε) and compares them with analytical gradients; a relative error below 1e-5 is considered correct, while an error above 1e-3 almost certainly indicates a bug.66% similarUnverifiedAt fp16 precision, each value occupies 2 bytes.65% similarUnverifiedFSRS-6 is the sixth-generation iteration of the FSRS series, and its accuracy in predicting forgetting curves in benchmark tests is significantly better than previous generations.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/494667API
curl https://kongchang.com/api/v1/knowledge/claims/494667MCP
get_claim(id=494667)