Edition of Thursday 12 November 2026

Tang, a compiler that derives GPU kernels and bounds how fast they can run

Comments on “Tang, a compiler that derives GPU kernels and bounds how fast they can run”

Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked

Municipal Robotics' Cam Pedersen describes tang, a Rust compiler that derives GPU kernels from a two-construct tensor IR (views and comprehensions) instead of using hand-written ones, including flash-style attention forward and backward. Reported results are mixed: 1.2× faster than PyTorch at sequence length 512 but slower at 128, and Qwen3-8B decode tied with llama.cpp at short context. Derived attention trails FlashAttention-3, and a Lean-proved traffic bound gives a time floor that is loose on the H200. The figures are the author's own and the claim check found no independent source.

Read the original at muni.works. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Why it ranked

No. 6 of 106 with a score of 0.71. The panel rated it highest on technical depth (0.98) and new information (0.92), lowest on audience breadth (0.59). Placed No. 6, No. 10 by score alone: the placement terms moved other stories past it.

AI judgeJudge 2claude-haiku-5-5
Substantive compiler/GPU kernel research with detailed benchmarks and proofs; strong technical depth and new information, though claims are partly self-reported and the paper is long and dense.
AI judgeJudge 3gpt-5.6-luna
A deeply technical compiler research report with reproducible implementation details, formal proofs, multi-backend evaluation, and candid limitations. It offers meaningful practical ideas for GPU and ML systems builders, though adoption and generality remain unproven.

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

0.71total under rubric v2.3

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.70
  • Reader interest0.84
  • Audience breadth0.59
  • Practical relevance0.66
  • Substantiated0.90
  • Technical depth0.98
  • New information0.92
  • Counts against: Promotional0.05
  • Counts against: Overclaim0.30

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.710.600.800.700.20+0.24+0.169
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.740.850.930.840.19+0.26+0.218
Audience breadthHow many people who build technology does it concern?0.550.600.620.590.07+0.13+0.077
Practical relevanceHow useful is it to someone who builds software or tech products?0.620.550.820.660.27+0.22+0.146
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.0.980.800.920.900.18+0.05+0.045
Technical depthHow technically substantive is it?0.990.950.990.980.04+0.05+0.049
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.890.900.980.920.09+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.0.060.050.030.050.03−0.19−0.009
OverclaimDo the headline or conclusions claim more than the body supports?0.400.300.200.300.20−0.11−0.033
Sum of contributions+0.708
Claim-check penalty0.000
Total, as published0.708

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.

Ranked #6 on Rubric, 12 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-12/6/"><img src="https://rubric.hrih.si/badge/2026-11-12/6.svg" alt="Ranked #6 on Rubric, 12 Nov 2026" width="238" height="28"></a>

Claim check

Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. No source givenBenchmark result

    On an idle RTX 3090, tang's derived transformer block forward pass at sequence length 512 is 1.2× faster than PyTorch captured in a CUDA graph with TF32.

    The article gives no link or citation for it.

    The article says

    “On an idle RTX 3090 a derived transformer block’s forward pass at sequence length 512 is 1.2× faster than PyTorch captured in a CUDA graph with TF32”

    The source says

    No source to compare with.

  2. No source givenBenchmark result

    Qwen3-8B decodes as a single IR graph at 148.7 tok/s, tied with llama.cpp at short context and at 92% of it at 4k.

    The article gives no link or citation for it.

    The article says

    “Qwen3-8B decodes as a single IR graph at 148.7 tok/s, tied with llama.cpp at short context and at 92% of it at 4k”

    The source says

    No source to compare with.

  3. No source givenBenchmark result

    With equal work, the derived MoE expert layer is 18% slower than the engine's hand kernels at batch 1 and 7% slower at batch 4, ties at 8 and is 5% faster at 64.

    The article gives no link or citation for it.

    The article says

    “the derived MoE expert layer is 18% slower than the engine’s hand kernels at batch 1 and 7% slower at batch 4, ties at 8 and is 5% faster at…”

    The source says

    No source to compare with.

  4. No source givenNumber

    Derived head-dim-256 attention and gated-delta scan cut the engine's time to first token at 44k tokens from 8.33 s to 5.54 s, but vLLM reaches first token about 12× faster than the engine.

    The article gives no link or citation for it.

    The article says

    “cut the engine’s own time to first token at 44k tokens from 8.33 s to 5.54 s, but vLLM on the FP8 checkpoint reaches first token about 12× faster than…”

    The source says

    No source to compare with.

  5. No source givenNumber

    The time floor never exceeds measured time on the 24 GEMV-class and epilogue kernels validated on two GPUs, reaching 0.76–0.93 of measured time on large RTX 3090 kernels and 0.05–0.66 (median 0.38) on the H200.

    The article gives no link or citation for it.

    The article says

    “On the RTX 3090 it reaches 0.76–0.93 of measured time on the large kernels. On the H200 it reaches only 0.05–0.66 (median 0.38”

    The source says

    No source to compare with.

  6. No source givenBenchmark result

    The derived attention runs at 0.43–0.54 of FlashAttention-3's speed forward and 0.22–0.40 backward.

    The article gives no link or citation for it.

    The article says

    “the derived attention runs at 0.43–0.54 of FlashAttention-3′s speed forward and 0.22–0.40 backward”

    The source says

    No source to compare with.

  7. No source givenNumber

    Inside the model, a derived decode step takes 5–12% less time at 16 to 48 concurrent streams on one H200, though the derived path misses the 99% agreement gate at 98.93%.

    The article gives no link or citation for it.

    The article says

    “a decode step takes 5–12% less time at 16 to 48 concurrent streams over five interleaved rounds on one H200”

    The source says

    No source to compare with.

  8. No source givenFirst, only or fastest

    Weak duality of the traffic linear program and the soundness of the core rewrites are proved in Lean.

    The article gives no link or citation for it.

    The article says

    “we prove weak duality in Lean for arrays indexed by subsets of the loop coordinates, along with the soundness of the core rewrites”

    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

Discussion on Hacker News

Recorded for context only, checked 12 Nov 2026, 05:47 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Comments

Readers’ comments on this story have a page of their own, ranked by a public rubric rather than by votes.

Comments on “Tang, a compiler that derives GPU kernels and bounds how fast they can run”

More on Rubric

From the last 7 editions, chosen by shared topic, score and date.

  1. Score 0.75

    llama.cpp v1.0