Revision History
NVIDIA H100 · 14 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: Corrected the H800's introduction date (March 2023, not November 2022), the H100 PCIe memory specs (the 94 GB HBM3 part is the H100 NVL), the GH100 TPC count, the DGX H100 chassis size (8U), the quarter in which GB300 overtook GB200, the FlashAttention-3 FP8 utilization figure and authors, and NVIDIA's actual $4.5 billion H20 charge; removed unsupported claims that GPT-4 and Gemini were trained on H100s, that Stargate Abilene used H100s, and an unsourced used-server price (plus 31 minor fixes)
Version 14 · Nov 1, 2026, 11:09 AM
- Unlink "scale AI" in a compound like "large-scale AI": it pointed to the company Scale AI
Version 11 · Oct 15, 2026, 05:10 PM
- Added an identifying image with caption and image credits
Version 10 · Oct 2, 2026, 08:17 PM
- Correction: fixed H100 SXM5 and PCIe dense versus sparsity-assisted Tensor Core throughput, corrected the DGX H100 FP8 aggregate, removed two ambiguously labeled cross-generation throughput rows, and refreshed the two supporting NVIDIA citation targets.
Version 9 · Jul 28, 2026, 07:31 AM
- Added 6 contextual internal links
Version 8 · Jul 23, 2026, 03:28 PM