A Megakernel per Model
Comments on “A Megakernel per Model”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Xingyu Dang reports that Claude coding agents wrote a single-GPU-program "megakernel" for each of eight open models, from 8B to 594B parameters. At batch size 1, with the reference data flow, the engines run 1.13–1.77× faster than the fastest of vLLM, SGLang or Cohere's own megakernel; allowing lossless storage changes raises this to 2.05×. On a production trace with Inkling-2-Small on 4×B200, 99–100% of requests met their latency targets versus 82–85% for SGLang. Gains shrink at batch size 8, and the post's own figures cover specific models and hardware.
Why it ranked
No. 6 of 140 with a score of 0.73. The panel rated it highest on technical depth (0.96) and new information (0.93), lowest on audience breadth (0.60). Placed No. 6, No. 8 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Detailed, well-measured engineering writeup on LLM inference megakernels with reproducible benchmarks and a public repo; technically deep and practically relevant to inference builders, though niche in audience and somewhat self-promotional in framing.
- AI judgeJudge 3gpt-5.6-luna
- A highly technical, reproducible study of agent-generated GPU inference engines, with extensive benchmarks, validation methodology, source artifacts, and explicit limitations. It offers unusually actionable performance and serving insights, though scope is model- and GPU-specific.
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
0.73total under rubric v2.3
- Significance0.72
- Reader interest0.86
- Audience breadth0.60
- Practical relevance0.74
- Substantiated0.90
- Technical depth0.96
- New information0.93
- Counts against: Promotional0.09
- Counts against: Overclaim0.24
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? | 0.68 | 0.60 | 0.88 | 0.72 | 0.28 | +0.24 | +0.172 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.78 | 0.85 | 0.95 | 0.86 | 0.17 | +0.26 | +0.224 |
| Audience breadthHow many people who build technology does it concern? | 0.54 | 0.55 | 0.70 | 0.60 | 0.16 | +0.13 | +0.077 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.66 | 0.70 | 0.86 | 0.74 | 0.20 | +0.22 | +0.163 |
| 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. | 0.98 | 0.80 | 0.93 | 0.90 | 0.18 | +0.05 | +0.045 |
| Technical depthHow technically substantive is it? | 0.99 | 0.90 | 0.99 | 0.96 | 0.09 | +0.05 | +0.048 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.89 | 0.90 | 0.99 | 0.93 | 0.10 | +0.05 | +0.046 |
| 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. | 0.06 | 0.15 | 0.05 | 0.09 | 0.10 | −0.19 | −0.016 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.28 | 0.25 | 0.20 | 0.24 | 0.08 | −0.11 | −0.027 |
| Sum of contributions | +0.732 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.732 | ||||||
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/2026-11-20/6/"><img src="https://rubric.hrih.si/badge/2026-11-20/6.svg" alt="Ranked #6 on Rubric, 20 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 3 matches source · 4 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceBenchmark result
With the same data flow as the reference implementation (track A), the megakernel engines are 1.13–1.77× as fast as the fastest general engine with one sequence, across all eight models from 8 B to 594 B parameters.
The cited source says the same thing.
The article says “the engines are 1.13–1.77× as fast as the fastest general engine with one sequence, on all eight models from 8 B to 594 B parameters”
The source says “megakernel, track A (s12 s0047) | 1.874 | 1.872–1.875 | 3.388 | 3.387–3.389 | 1.77× | 1.45×”
Source cited by the article: dangxingyu.github.io/megakernel/full.html
Matches sourceBenchmark result
Under track B, where model storage may change but its numerical function may not, the speedup factor reaches 2.05×.
The cited source says the same thing.
The article says “When the storage of the model may change but its function may not, numerically (track B), the factor reaches 2.05×”
The source says “megakernel, track B (s12 s0050) | 1.615 | 1.615–1.616 | 3.388 | 3.387–3.389 | 2.05× | 1.46×”
Source cited by the article: dangxingyu.github.io/megakernel/full.html
Matches sourceBenchmark result
Serving a production trace on Inkling-2-Small and 4 × B200 at 1 to 8 requests per second, 99–100% of requests meet service levels versus 82–85% for SGLang, and the longest stall is 0.08–0.15 s versus 2.3–4.4 s.
The cited source says the same thing.
The article says “99–100% of requests stay within the service levels against 82–85% for SGLang, and the longest stall is 0.08–0.15 s against 2.3–4.4 s”
The source says “megakernel, serving, track A | 238 | 100.0 | 0.19 | 3.98 | 27.9 | 69 | 75 | 166 | 2 | rel-inkon2-replay-r1”
Source cited by the article: dangxingyu.github.io/megakernel/full.html
No source givenNumber
On Qwen3-8B and an H100, reading the 15.1 GB of BF16 weights once at 3.35 TB/s takes 4.52 ms, and the track-A engine runs at 5.13 ms per token, 88% of that bound.
The article gives no link or citation for it.
The article says “the track-A engine runs at 5.13 ms per token, 88% of that bound”
The source says No source to compare with.
No source givenCausal conclusion
On Qwen3-8B, lossless 12-bit weight encoding decoded in registers cut weight traffic per token by a quarter.
The article gives no link or citation for it.
The article says “encoded each weight losslessly in 12 bits and decoded it in registers ahead of the tensor cores, cutting weight traffic per token by a quarter”
The source says No source to compare with.
Couldn’t checkNumber
The engines in the post took 19 runs, 458 agent-hours and 675 submissions.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “The engines in this post took 19 runs, 458 agent-hours and 675 submissions”
The source says No passage could be compared.
Source cited by the article: dangxingyu.github.io/megakernel/full.html
No source givenNumber
All 20 re-tested engines (all except the three Inkling-2-Small engines) pass every check of the current test suite.
The article gives no link or citation for it.
The article says “20 of 20 pass every check”
The source says No source to compare with.
No source givenFirst, only or fastest
Until now megakernels have been written by hand for a few models (Hazy Research's for Llama-1B, Cohere's for North Mini Code) or compiled for supported architectures (MPK).
The article gives no link or citation for it.
The article says “until now they have been written by hand for a few models (Hazy Research's for Llama-1B3, Cohere's for North Mini Code4) or compiled for a set of supported architectures (MPK5”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 20 Nov 2026, 05:47 UTC. It played no part in the ranking.