Soumith Chintala’s Post

Interesting flowchart Soumith Chintala. Another question is whether they can spell software, which will ultimately deliver the user experience from the quick benchmarking claims and HW spec.

Very nice viewpoint. Could you elaborate the claims of Groq LPU engine as a startup with the correlation in your chart:- "They want to do what the Nvidia does best, But do it better than Nvidia." Do you still have a laugh for Groq LPU or you think its a serious business.

I wonder how closely is this related to etched.ai 's current pitch. . . 🤔

No, I don’t agree with following points, 1. Team Competence 2. Model heterogeneity 3. Better than Nvidia The hardware startup’s have niche talent they know how to optimise the model for best performance. They don’t want to compete with Nvidia, there aim is to make models works on their platforms so that it can align with company’s goals and achieve maximum customer satisfaction. Making model compatible with different hardware chips is very complicated. It is not about results always, it is about making the same model work on different chips which may not have same specifications and configurations.

Forget startups, even the hyperscaler chips bullshit their way around. For the longest time you couldn't get a straight answer on convergence metrics for popular models, the only answer they'd give you in the Flops and Cost!

Thanks for the diagram, makes sense - and to add to it, hardware startups have a lot of vectors to optimize already - power, area, speed, system design, tapeout, system integration; not to talk about the human factor, a hardware startup requires highly specialized/trained/experienced people - they use a set of benchmarks (SoecINT, CoreMark, MLPerf) to come up with a microarchitecture / system they can claim is “better” (for some definition of better); and like anything that is in its infancy, they do not want to loose a sale for some silly performance bug. I personally consider [hardware] startups as still maturing after 2-5 years after inception (depending on the system they are designing).

Too true. The only other question is if they've lowered accuracy targets or not in order to improve performance.

hi. how are you? I hope you are very well. Iam gald to speak you. I am glad to talk to you Do you live in India? what are you doing? 😃

You want a chip faster than NVIDIA, yet general enough to run wacky models not capturable by onnx graphs. All due respect, I think this is physically impossible.

I think very few people in the world would be able to make this assessment. You are very knowledgeable.