GPU MODE linear-algebra kernel audit, 2026: results record
What the record is
The record holds the results of our independent audit of GPU MODE's 2026 competitions for fast GPU linear-algebra kernels. The audit covers the five top-ranked entries on each of four leaderboards, for batched QR in two rounds, batched Cholesky and batched symmetric eigendecomposition.
The organizers checked each entry's accuracy on a fixed set of cases and ranked the entries that passed by speed. The audit ran each entry, unmodified, through the organizers' own checker on further legal inputs, on the GPU model the competitions graded on.
This page describes version 1.0.2 of the record, git tag v1.0.2, released on 2026-11-11.
Our paper describes the audit and cites this record.
Terms
- Scored cases. Scored cases are the organizers' own case specifications from the task files in GPU MODE's
reference-kernelsrepository, run at their listed seeds. For QR they are the benchmark cases only, 7 in round 1 and 12 in round 2. - Legal input. A legal input is one the task allows. For round-2 QR the task statement bounds neither the conditioning nor the structure of the float32 matrices.
- Search. The search is the audit's own set of further inputs.
- Leaderboard number. A leaderboard number is GPU MODE's number for the leaderboard of one competition.
- Vendor routine. The vendor routine is the PyTorch call
torch.geqrf,torch.linalg.cholesky_exortorch.linalg.eighon the GPU. - Failure, for QR. A QR failure is one searched evaluation whose output was not vendor-identical and that the round-2 checker rejected. An output is vendor-identical when it equals, bit for bit, the vendor routine's output on the same input.
- Failure, for Cholesky and eigendecomposition. A search failure is attributable to the entry when the organizers' checker fails it, the vendor routine passes, and the input passed validation.
- Reproduced. For Cholesky and eigendecomposition, five failing search inputs per entry became candidates, each rerun three times. A candidate is reproduced when all three reruns fail while the vendor routine passes.
Summary
Each count in the last three columns is a number of entries.
| Round | Leaderboard number | Passed the scored cases of its round | Failed in the search | Failure reproduced in reruns |
|---|---|---|---|---|
| QR, round 1 | 773 | 5 of 5 | 5 of 5 | not rerun |
| QR, round 2 | 774 | 5 of 5 | 5 of 5 | not rerun |
| Cholesky | 776 | 5 of 5 | 5 of 5 | 4 of 5 |
| Eigendecomposition | 775 | 5 of 5 | 3 of 5 | 3 of 5 |
The record's README summarizes the table in this sentence.
"All twenty entries passed the scored cases of their own round, and eighteen failed the same checker on other legal inputs on which the vendor routine (the PyTorch call, backed by NVIDIA's libraries) passed."
What the record contains, and what it does not
The record contains these items.
- The verdicts.
- A description of every recorded evaluation, enough to regenerate the input of every recorded QR evaluation.
- The hashes of the entries, checkers and environments.
- Since version 1.0.2, 48 scored runs of the round-2 QR baseline, GPU MODE's starter submission, found in the organizers' dataset, which give its speed.
- Scripts that rebuild every CSV of the audit's results from the raw files and recompute every table in the README.
It does not contain these items.
- The code of any entry. The organizers publish the code of the round-2 QR, Cholesky and eigendecomposition entries, and the round-1 QR code is not public.
- The software that searched for the failing inputs.
- The generators of the Cholesky and eigendecomposition search inputs.
What can be replayed
Every recorded QR input is determined by its row in qr/sweep.csv or qr/scored.csv. Round-2 QR entries can be fetched from the organizers' public dataset and replayed with replay/replay_qr.py, which prints the verdict under both checkers. A replay needs Linux, an NVIDIA B200, a CUDA toolkit with nvcc 13, a C++ compiler, and the pinned Python packages. The README warns that its second example sits just above its bisection threshold, so a small difference in environment can flip its verdict.
Round-1 QR entries are not in the organizers' dataset, so their verdicts cannot be replayed from public sources unless the organizers release the round-1 code. The Cholesky and eigendecomposition search inputs came from the audit's own generators, which are not part of the record. Their verdicts can be inspected, but the inputs cannot be regenerated from this record alone. The scored cases of these two problems are the organizers' and are reproducible with their code.
License
The audit's own results are copyright 2026 Maksim Melnichenko and released under CC BY 4.0, in the file LICENSE. The code in scripts/, replay/ and qr/randsvd.py is released under the MIT License, in the file LICENSE-CODE. Some of the data comes from GPU MODE and stays under GPU MODE's terms, namely the scored-case specifications, the leaderboard metadata and the starter runs. The record copies no entry code.
How to cite
Cite the version you used.
| Field | Value |
|---|---|
| Title | GPU MODE linear-algebra kernel audit, 2026: results record |
| Author | Maksim Melnichenko |
| Version | 1.0.2 |
| Date released | 2026-11-11 |
| Repository | https://github.com/the-numerical-proving-ground/gpumode-kernel-audit-2026 |
When you use the GPU MODE material, also link the GPU MODE dataset and cite KernelBot, as the dataset card requests.