From 2e102b9115682119947b3678d42fb5cc1bca0817 Mon Sep 17 00:00:00 2001 From: Mark Saroufim Date: Mon, 13 Jul 2026 17:16:38 -0700 Subject: [PATCH] Add Cholesky to linear algebra post --- kernelboard/static/news/2026-06-12-linear-algebra-kernels.md | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/kernelboard/static/news/2026-06-12-linear-algebra-kernels.md b/kernelboard/static/news/2026-06-12-linear-algebra-kernels.md index ee972d7..ff17c9b 100644 --- a/kernelboard/static/news/2026-06-12-linear-algebra-kernels.md +++ b/kernelboard/static/news/2026-06-12-linear-algebra-kernels.md @@ -36,6 +36,10 @@ The main correctness footgun is that eigenvectors are not unique. If `q` is an e Instead the checker looks at the mathematical invariants: `Q.T @ Q ~= I`, `A @ Q ~= Q @ diag(L)`, `Q @ diag(L) @ Q.T ~= A`, sorted eigenvalues, finite FP32 outputs, and the expected shapes/devices. Like QR, the tolerances are intentionally residual-based and scaled by `n * eps32` so approximate or low-bit internal strategies should work. +## Cholesky + +The third problem is batched FP32 Cholesky factorization: given a symmetric positive-definite matrix `A`, return a lower-triangular `L` such that `A = L @ L.T`. The 15 B200 benchmarks span thousands of small matrices, paired low- and high-batch cases from `512 x 512` through `4096 x 4096`, and single matrices up to `32768 x 32768`. Correctness is property-based, and the problem concludes July 30, 2026. + ## Prize We'll be using a simple scoring system: if any of your submissions are in the top 3 of any problem then you'll be recognized as a winner.