Hacker News Reader

Comment by 7e | Hacker News Reader

100Gbps Ethernet cards? The world has moved way past that for training. Their accelerator stack must be really slow if this is good enough for them.

View on HN

It doesn't say they can't aggregate links. I wouldn't say the world has moved way past that yet, but Tesla probably doesn't want to be dependant, like everyone else but Apple, on Nvidia (InfiniBand).

> like everyone else but Apple, on Nvidia (InfiniBand)

Google also does not depend on NVIDIA, thx to TPUs. Rents NVIDIA GPUs to external customers - sure, it's a nice side business, but internally TPUs are king and there's no dependency on NVIDIA for that.

Likewise Microsoft and Meta have developed in house AI chips, but I don't know what fraction of their AI workloads run on them.

Not a meaningful amount :). Their “AI chips” are, for now, marketing.

From what I know, Meta AI chips are used in production today, but are made for their recommendations tasks which is a very different IA than GPTs and LLMs for which they still rely on GPUs.

TPUs are essentially garbage compared to NVIDIA hardware. TPUs are king of nothing, but a primary ingredient in Kool-Aid

This is not true at all.

It's very hard to respond to a comment like that, since there's no specifics, just a plain disagreement.

On my side, I would like to point that the today HN thread ([1]) that discusses a paper GameNGen ([2]) that runs Doom with diffusion models was trained on TPUs.

I don't see a dependency on NVIDIA there.

If there's a more specific rebuttal to my original statement, please, don't hesitate to state it.

1. https://news.ycombinator.com/item?id=41375548

2. https://gamengen.github.io/

3. https://arxiv.org/abs/2408.14837

> Google also does not depend on NVIDIA,

Deepmind says otherwise. training is most likley all on NVIDIA still. Same for Apple.

The difference is, nobody knows for sure with Apple, because they are a secrecy cult.

You can get 400gbit ethernet from half a dozen vendors.

Reply on news.ycombinator.com