/Tech3h ago

GPU MODE, co-founded by PyTorch maintainer Mark Saroufim, launches a QR decomposition benchmark and leaderboard on Nvidia B200 hardware

Story Overview

GPU MODE has refreshed its batched QR decomposition benchmark as qr_v2 and opened a public leaderboard running on Nvidia B200 hardware. The update adds stricter checks for degenerate matrices, random seeds, and residual validation while keeping submissions open to anyone via CLI or browser. A hand-written CUDA kernel currently sits at the top with a 1,999-microsecond time.

8751158.1K

#102

Original post

Thien Tran@gaunernst

How I feel beating Codex at QR32 with a hand-written kernel (Codex has moved on to other shapes and I'm still at the bottom of the leaderboard)

Mark Saroufim@marksaroufim

We've released the QR problem, a more robust qr_v2 with a fresh leaderboard so please resubmit!

Thank you to @blelbach, @myainotez and @nikhilbarhate99 for sharing feedback. Sorry if I missed anyone!

I considered automatically backfilling all submissions but the rankings do change quite a bit so I figured a refresh would be better.

Changelog * Fail submissions if they fail when we change random seeds * Add nasty correctness cases with more degenerate inputs in mixed batches * Recheck correctness when doing perf testing to avoid Volkswagen cheat * Reject Nan/Inf residuals * Validate each matrix factorization residual, since averaging was hiding bad matrices * Old QR is still open so folks can't see submissions but you can't submit anything to it

Wontfix * Stream hacking is still banned via very blunt ban of the word "stream" we don't have a good solution for this * CUDA graphs are allowed but not particularly interesting to us

Best submissions so far if I resubmit their solutions are

7:13 AM · Jun 16, 2026 · 5.3K Views

Developer Impact

Hand-written kernels still beat Triton entries

The current leader, badelsteinlelbach, posted 1999 μs using a plain submission.py file. The next two spots use Triton-based code and trail by more than 600 μs, showing that custom CUDA work can still deliver measurable speed gains on this workload.

Open Question

Stricter validation leaves room for edge cases

The refresh closed the prior leaderboard and introduced new test matrices, but the exact scoring formula, full matrix list, and any stream-hacking rules remain unspecified in the announcement. It is unclear whether future submissions will face additional hidden constraints.

Sentiment

Many users praised the handwritten QR kernel beating Codex on the leaderboard as artistic and a fun learning experience, while some urged not to slow Codex down.

Pos

75.0%

Neg

25.0%

4 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS2.9KBOOKMARKS10LIKES23REPLIES4

rohan anil@_arohan_

Leaderboard today for QR 🚀

Cuda Colonel has taken a sizeable lead

——

This benchmark is on various shapes and distribution of matrices.

Some of the matrices is either zero, tiny, or numerically low-rank: rankdef, clustered, nearcollinear

Competitors would do well by tailoring their solver to various paths

Not all operations in fp32. Reflectors are hard particularly but not others.

Results on leaderboard so far are already exceeding my expectations.

Mark Saroufim@marksaroufim

We've released the QR problem, a more robust qr_v2 with a fresh leaderboard so please resubmit!

Thank you to @blelbach, @myainotez and @nikhilbarhate99 for sharing feedback. Sorry if I missed anyone!

I considered automatically backfilling all submissions but the rankings do change quite a bit so I figured a refresh would be better.

Wontfix * Stream hacking is still banned via very blunt ban of the word "stream" we don't have a good solution for this * CUDA graphs are allowed but not particularly interesting to us

Best submissions so far if I resubmit their solutions are

3h2.9K2310

sankalp@dejavucoder

@_arohan_ i find it crazy how far one can come by agentmaxxing. i didn't know about qr decomposition 2 days back😂

2h592

rohan anil@_arohan_

@dejavucoder Did you learn about it already? Or is it worth for me making a tutorial post about it?

2h392

Thien Tran@gaunernst

my favourite incantation was attempting this challenge with CuteDSL, but it seems nvcc gives better codegen for this kind of problem

3h1052

sankalp@dejavucoder

@_arohan_ i watched a couple of videos and brainstormed with claude.

re:tutorial, i think people interested in infernece stuff would still appreciate a tutorial. would be worth it if it's not high effort for you

2h362

Sachin@sachdh

@_arohan_ This is fun competition.

I am learning quite a bit while vide kerneling

3h352

Sinatras@myainotez

@gaunernst Its also a form of art when you do this

3h342

sankalp@dejavucoder

@_arohan_ stuck at 3956 for a while now

3h35

Mateusz Kelner@MKelner

@_arohan_ 4196, I’m coming for you @blelbach

2h91

Simon V@Simon_Vt

@gaunernst don't slow codex down 😡

2h22

Mateusz Kelner@MKelner

@_arohan_ Every time I think I have a shot at competing the top 3 just pulls away. Insane crowd! Currently sitting at 4615

3h5

Zeus@ZeusSkils

@_arohan_ u can run a few hundred cublas masks and find spots to go pro with kernels that skip zero tiles. geometry matters more than any flavor of linear algebra here

2h2