Show HN: A web-based replacement for Nvidia's CUDA occupancy spreadsheet (toolbelt.widgita.xyz)

1 points by fairlight1337 ↗ HN
NVIDIA still ships their occupancy calculator as an Excel file from the SM 7.x era. I wanted something I could link a colleague to, that supports newer architectures (Hopper, Ada, Blackwell) and shows the sweep across block sizes inline.

It's pure client-side JS — no backend, no tracking, no signup. The math lives in formulas.js so you can sanity-check it against the CUDA Programming Guide. Architectures and their per-SM limits are in presets.js.

Things I'd love feedback on:

* Are the per-arch limits correct for [arch you're least sure about]?

* The "compare two configs side-by-side" view - is the diff readable?

* Anything you wish the Excel version did that this doesn't?

Link: https://toolbelt.widgita.xyz/cuda-occupancy-calculator/

(There are a few other small tools at the root, but this one is the reason for the post.)

1 comment

[ 3.1 ms ] story [ 12.7 ms ] thread
How are you handling register allocation granularity across architectures? Small changes in regs per thread can shift the limiter in non-obvious ways