Internals of FP8 MMA on Hopper
Comments on “Internals of FP8 MMA on Hopper”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
A blog post by Pieter Tillet's author (unnamed in the text) explains why FP8 matrix multiply on NVIDIA Hopper GPUs needs accumulator promotion. The post attributes Hopper's accuracy loss to a truncating alignment cutoff and a rounding format that zeroes the lowest 10 significand bits, then shows that periodically promoting…
Why it ranked
No. 22 of 107 with a score of 0.61. The panel rated it highest on technical depth (0.96) and substantiated (0.88), lowest on audience breadth (0.44). Placed No. 22, No. 47 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Deep, well-evidenced technical write-up on FP8 accumulation numerics on Hopper GPUs, with reproducible code and comparisons; relevant mainly to GPU kernel and ML infrastructure builders, not broad tech audiences.
- AI judgeJudge 3gpt-5.6-luna
- Detailed, reproducible technical analysis of Hopper FP8 tensor-core behavior, with simulator code, GPU kernel validation, numerical examples, and practical mitigation guidance. Its audience is narrower, but the evidence and practitioner value are strong.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.61total under rubric v2.3
- Significance0.55
- Reader interest0.67
- Audience breadth0.44
- Practical relevance0.65
- Substantiated0.88
- Technical depth0.96
- New information0.81
- Counts against: Promotional0.06
- Counts against: Overclaim0.17
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.52 | 0.40 | 0.72 | 0.55 | 0.32 | +0.24 | +0.131 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.50 | 0.60 | 0.90 | 0.67 | 0.40 | +0.26 | +0.173 |
| Audience breadthHow many people who build technology does it concern? | 0.39 | 0.30 | 0.62 | 0.44 | 0.32 | +0.13 | +0.057 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.58 | 0.55 | 0.82 | 0.65 | 0.27 | +0.22 | +0.143 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.95 | 0.80 | 0.88 | 0.88 | 0.15 | +0.05 | +0.044 |
| Technical depthHow technically substantive is it? | 1.00 | 0.90 | 0.97 | 0.96 | 0.10 | +0.05 | +0.048 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.90 | 0.60 | 0.92 | 0.81 | 0.32 | +0.05 | +0.040 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.05 | 0.05 | 0.08 | 0.06 | 0.03 | −0.19 | −0.011 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.17 | 0.15 | 0.18 | 0.17 | 0.03 | −0.11 | −0.018 |
| Sum of contributions | +0.607 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.607 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-03/22/"><img src="https://rubric.hrih.si/badge/2026-11-03/22.svg" alt="Ranked #22 on Rubric, 3 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 2 matches source · 6 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceNumber
Hopper's FP8 MMA rounding format writes zeros into the lowest 10 bits of the F32 significand, which could hold 23 bits.
The cited source says the same thing.
The article says “this format writes all zeros into the lowest 10 bits of the F32 significand which could hold potentially 23 bits”
The source says “QMMA instructions on NVIDIA Ada Lovelace GPUs and the QGMMA instructions on NVIDIA Hopper GPUs employ only F=13 fractional bits for fused summation within the T-FDPA operation and use \rho=\text{RZ}_{\text{E8M13”
Source cited by the article: arxiv.org/abs/2511.10909
No source givenBenchmark result
The MMA-Sim bitwise model matches real GPU output for the wgmma m64n8k32 E4M3 instruction on an H100 (and an H200), with bitwise-equal results.
The article gives no link or citation for it.
The article says “MMA-Sim 0x43809800 GPU 0x43809800 bitwise equal: True”
The source says No source to compare with.
No source givenNumber
DeepGEMM uses promotion in steps of 128 K, which corresponds to 4 WGMMA ops.
The article gives no link or citation for it.
The article says “DeepGEMM uses the promotion in steps of 128 K, which means 4 WGMMA ops”
The source says No source to compare with.
No source givenNumber
The Hopper FP8 instruction width is 32, so one instruction performs a 32-term dot product fused with the existing accumulator.
The article gives no link or citation for it.
The article says “the instruction width of the FP8 MMA is 32, so a dot product of 32 terms is essentially performed with one instruction”
The source says No source to compare with.
Matches sourceVersion
Blackwell fixes the Hopper limitation, using a cutoff at F = 25 and F32 accumulator rounding toward zero.
The cited source says the same thing.
The article says “is fixed on Blackwell, where we have both the cutoff at F = 25 and F32 accumulator rounding toward zero”
The source says “For FP8 formats, F increases to 25 , and \rho changes from \text{RZ}_{\text{E8M13}} to \text{RZ}_{\text{FP32}} for FP32 outputs”
Source cited by the article: arxiv.org/abs/2511.10909
No source givenVersion
Hopper does not support hardware block scaling, while Blackwell does.
The article gives no link or citation for it.
The article says “on Hopper, block scaling via the hardware is not supported, so we have to do it manually”
The source says No source to compare with.
No source givenNumber
At K = 16384 on positive data, B200 chained has relative error 6.05e-6, while H200 has 2.49e-1 chained and 4.96e-4 promoted.
The article gives no link or citation for it.
The article says “the chained B200 has a relative error of 6.05e-6, while the H200 gives 2.49e-1 chained and 4.96e-4 promoted”
The source says No source to compare with.
No source givenGeneralization
Blackwell without promotion is two orders of magnitude more accurate than Hopper with DeepGEMM-style promotion.
The article gives no link or citation for it.
The article says “Blackwell without any promotion is two orders of magnitude more accurate than Hopper with DeepGEMM-style promotion”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 3 Nov 2026, 05:47 UTC. It played no part in the ranking.