Revision History
FP8 · 7 revisions
Sizes are character counts of the article source. The signed number is the change from the previous revision.
Recent edit summaries
Detailed summaries recorded by editors. Generic maintenance summaries are omitted here; every recorded revision remains below. Dates describe the edit, not necessarily the event it covers.
- Correction: re-cited the H100 FP8 throughput figures to NVIDIA's final product specifications (the earlier source gave preliminary numbers), updated the moved vLLM FP8 documentation link, fixed a misquoted DeepSeek-V3 line and the scope of its 0.25% loss-error result, replaced a dead Transformer Engine citation; Update: added FP8 conversion/saturation rules, FNUZ details, hardware table, scaling-strategy section with Transformer Engine recipes, layers kept in higher precision (incl. NVIDIA DGX Spark online FP8 LM head example), static vs dynamic quantization, FP8 KV cache and attention, E8M0 scale rounding
Version 7 · Oct 29, 2026, 09:33 PM
- Added an explanatory diagram with caption and image credits
Version 6 · Oct 4, 2026, 06:02 PM
- Add Meta GEM (August 3, 2026) as a production-scale MXFP8 training deployment in the MX-formats section, with new reference [24]
Version 5 · Aug 5, 2026, 11:48 AM
- Added 6 contextual internal links
Version 4 · Jul 23, 2026, 04:05 PM