[amelius]

But AI depends on a small number of tensor operators, primitives which can be relatively easily implemented by competitors, so compute is very close to being a commodity when it comes to AI.

A company like Cerebras (founded in 2015) proves that this is true.

The moat is not in computer architecture. I'd say the real moat is in semiconductor fabrication.

[bilbo0s]

which can be relatively easily implemented by competitors

Oh my.

Please people, try to think back to your engineering classes. Remember the project where you worked with a group to design a processor? I do. Worst semester of my life. (Screw whoever even came up with that damn real analysis math class.) And here's the kicker, I know I'll be dating myself here, but all I had to do for my part was tape it out. Still sucked.

Not sure I'd call the necessary processor design work here "relatively easy"? Even for highly experienced, extremely bright people, this is not "relatively easy".

Far more easy to make the software a commodity. Believe me.

[amelius]

To be totally honest, the only thing I can distill from this is that perhaps you should have picked an education in CS instead of EE.

I mean this is like saying that a class for building compilers sucked. Still, companies make compilers, and they aren't all >$1B companies. In fact, hobbyists make compilers.

[bilbo0s]

I did study CS as well.

That you are comparing designing and writing a compiler with designing and manufacturing a neural processor is only testimony to the futility of my attempt to impress on everyone the difference. So I'll take my leave.

You have a good day sir or ma'am.

[amelius]

But I'm actually saying that manufacturing is the hard part ...

[sounds]

Have you ever tried to run a model from huggingface on an AMD GPU?

Semiconductor fabrication is a high risk business.

Nvidia invested heavily in CUDA and out-competed AMD (and Intel). They are working hard to keep their edge in developer mindshare, while chasing hardware profits at the same time.

[phkahler]

>> Have you ever tried to run a model from huggingface on an AMD GPU?

Yes. I'd never touched any of that stuff and then one day decided to give it a shot. Some how-to told me how to run something on Linux which had a choice of a few different LLMs. I picked one of the small ones (under 10B) and had it running on my AMD APU inside of 15 minutes. The weights were IIRC downloaded from huggingface. The wrapper was not. Anyway, what's the problem?

BTW that convinced me that small LLMs are basically worthless. IMA need to go bigger next time. BTW my "old" 5700G has 64GB of RAM, next build I'll go at least double that.

[amelius]

> Have you ever tried to run a model from huggingface on an AMD GPU?

No, but seeing how easily they run on Apple hardware, I don't understand your point, to be honest.

[ants_everywhere]

> The moat is not in computer architecture. I'd say the real moat is in semiconductor fabrication.

In the longer run, anything that is very capital intensive, affects entire industries, and can be improved with large amounts of simulation will not be a moat for long. That's because you can increasingly use AI to explore the design space.

Compute not a commodity yet but may be in a few years. Semiconductor fab will take longer, but I wouldn't be surprised to see parts of the fabrication process democratized in a few years.

Physical commodities like copper or oil can't be improved with simulation so they don't fall under this idea.

[recursivecaveat]

It's not like you can just stamp out a giant grid of flops and just go brrr. Getting utilization is difficult, and the closer you hew to Nvidia's tradeoffs the more you are going to come out unfavorably against a giant who's working with 10,000X your volume and decades of experience. Nvidia proprietary software is very highly embedded into everyone's stacks. The models undergo co-evolution with the hardware, so they are designed with its capabilities in mind.

It's like trying to take on UPS with some new, not quite drop-in logistics network. Theoretically its just a bunch of empty tubs shuffling around, but not so easy in practice. You have to be multiples better than the incumbent to be in contention. Keep in mind for the startups we don't really know who is setting money on fire running models in unprofitable configurations for revenue.