Hacker News new | ask | show | jobs
by jordiburgos 34 days ago
This is very helpful too: https://www.canirun.ai/
6 comments

Love that it defaults to the GPU being "NVIDIA GeForce 8800 GTX", a GPU released in 2006 with ~700MB of VRAM...

The estimates seems far off as well, took https://www.canirun.ai/model/gpt-oss-120b as an example, with a RTX Pro 6000 and every single number is off, and notably misses estimation for the most important quant for GPT-OSS, the MXFP4 variant.

The default for me was M1. I think it tries to guess using WebGPU.
> canirun.ai

I run dgx spark, and the results here are soooo incomplete for my platform that I can’t trust this site (for my usecase).

Yes, I really like this site too, but it's a bit outdated.

"39d ago" in AI time is like 1 year outdated info.

I don't think this has been the case for at least 39 days. The news is slowing down. The big headlines now, besides unverified marketing claims, are efficiency gains. Which are fantastic, but don't seem to be met with matching performance gains.
Every browser gives me a different result, I guess I can't blame the site for that. But it should perhaps mention which browser would be the most accurate.
Memory bandwidth is completely different on any browser from measured results on my M2 Pro machine. Weirdly, the estimated performance levels and even exact product name differ between Chrome and Firefox. Firefox calls it an M2 Pro and overshoots measured memory bandwidth by 40GB/S, Chrome calls it an "Apple M2 Pro" and overshoots by 80.
Doesn’t have qwen3 coder next. What else is it missing?
canirun doesn’t work well for MoE models. Qwen’s 35B A3B model works amazing on my M1 Max 32GB, but canirun thinks it barely runs.