Unverified50% confidenceBenchmarkExact time
FA4 MX8 achieves performance of 2.85 PF/s for the forward pass and 2 PF/s for the backward pass on typical LLM shapes.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/16/2026
First Seen
Valid until: 12/15/2026
Sources
Related Entities
Related Claims
UnverifiedMXFP8 has two variants: E4M3 (4-bit exponent, 3-bit mantissa) and E5M2 (5-bit exponent, 2-bit mantissa).65% similarUnverifiedLow-precision FlashAttention-4 (FA4 MX8) builds on FlashAttention-4 by fully migrating both forward and backward propagation to the MXFP8 microscaling low-precision format.63% similarUnverifiedFP8 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.62% similarUnverifiedNVIDIA's H100 GPU natively supports the FP8 format, with throughput approximately 2× that of FP16.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/927323API
curl https://kongchang.com/api/v1/knowledge/claims/927323MCP
get_claim(id=927323)