From 91a630becd671f45642ca20a49d0b6cf86086ac4 Mon Sep 17 00:00:00 2001 From: luciferlive112116 <291889058+luciferlive112116@users.noreply.github.com> Date: Wed, 15 Jul 2026 21:17:31 +0800 Subject: [PATCH] fix(docs): landing page still advertises the pre-8192 reference regime index/index.html named N = 12,000 as CCO's reference regime and shipped a copy-pasteable 'python -m eval --n 12000 --pairs 3' quickstart, but the engines and every other surface moved to 8192: strategy/matmul --n default is 8192, and README.md / BENCHMARKS.md both state 'Reference setup: 8192 x 8192'. The KPI beside it (3.5 TFLOP) was right only for the stale N -- 2*12000**3 = 3.456e12 -- so at the real default it must read 2*8192**3 = 1.0995e12 -> 1.1 TFLOP. The neighbouring 1,000 -> 2 GFLOP and 128,000 -> 4.2 PFLOP cards already use decimal N and stay as they are. Re-derive the three stale claims from the shipped default and add a CPU test that pins the page's advertised N, quickstart and KPI to strategy/matmul's --n default so the front page cannot drift away from the engines again. --- index/index.html | 8 +- tests/test_landing_page_reference_regime.py | 83 +++++++++++++++++++++ 2 files changed, 87 insertions(+), 4 deletions(-) create mode 100644 tests/test_landing_page_reference_regime.py diff --git a/index/index.html b/index/index.html index 24092bd..769d056 100644 --- a/index/index.html +++ b/index/index.html @@ -334,8 +334,8 @@
Trivial today — the N³ term means this is not where the real cost lives.
Full-rank data, deliberately the hard case — no structure to exploit for free.
Approximate and exact-but-cheaper strategies for C = A × B, scored on full-rank, low-rank, and decaying-spectrum regimes. Open for contributions today.
- +Clone the repo, install PyTorch, and reproduce the matmul reference scorecard on your own GPU — the same numbers a reviewer checks.
python -m eval --n 12000 --pairs 3python -m eval --n 8192 --pairs 3