[otabdeveloper4]

> we're building an enormous parallel computing infrastructure that's going to revolutionize scientific computing

We're doing an absolute shit job of it and basing our foundations on piles of sand. We're gonna have to throw it all away and do it right the second time if we ever decide to do something useful.

[smackeyacky]

I don’t think that matters. Out of the Dotcom rubble, we got a lot of useful ideas for tools even though the actual tools all got thrown away. It was the ideas and the “hey, look at what is possible” that survived i.e. Linux came of age properly and now the world runs on it.

AI will quickly pollute itself, but the infrastructure will survive. That most of it will be thrown away is perfectly ok, the ideas will live on.

[otabdeveloper4]

> but the infrastructure will survive

What infrastructure? Massively inefficient corporate rent-seeking attempts aren't infrastructure.

There is no AI "Linux" in 2024, all AI tools are closed-source and proprietary.

[__loam]

I'm not sure I agree with that and I'm one of the biggest ai pessimists here. The gpus getting created to do this shit aren't just gonna get thrown out (maybe sold in a fire sale), and they're pretty general purpose. Tons of science needs massive compute.

[otabdeveloper4]

Using GPUs for neural network inference is inefficient and makes no sense, unless there's some hype bubble propping it all up.

Maybe sometime in the future a commodity manufacturer of massively parallel CPUs for stupidly parallel tasks appears, but Nvidia ain't it.

[happymellon]

How many cores/threads would you need to challenge the GPU?

Ampere has 320 core chips as a dual socket, although a 4090 might still be cheaper.

[otabdeveloper4]

80% of the effort and cost of a GPU is various bullshit related to rendering DirectX games on Windows.