GPUs for Deep Learning in 2023 – An In-depth Analysis

Re:GPUs for Deep Learning in 2023 – An In-depth Analysis by nottorp on Thursday January 19, 2023 @07:00AM (#34439415)

What's very interesting to me if i read the "slowdown vs power limit" graph right, is you can get 74% of the performance for half the power consumption?

And since I know nothing about deep learning:

Can you do these neural networks on AMD GPUs at all, regardless of performance? Because the hive mind says AMD GPUs are better if you run Linux.

Can you do these neural networks on Linux at all?

[ Parent | on HN ]

Re:GPUs for Deep Learning in 2023 – An In-depth Analysis by paol on Thursday January 19, 2023 @07:35AM (#34439641)

> Can you do these neural networks on AMD GPUs at all, regardless of performance?

In principle yes, in practice it's not worth the bother. Scroll to the FAQ section at the end of the article, they cover the AMD situation pretty well.

> Because the hive mind says AMD GPUs are better if you run Linux.

As usual the hivemind just regurgitates memes without understanding.

> Can you do these neural networks on Linux at all?

Almost all NN work is done on Linux. Recent WSL2 developments have supposedly made Windows viable as well, but I don't know how popular it is.

[ Parent | on HN ]

Re:GPUs for Deep Learning in 2023 – An In-depth Analysis by nottorp on Thursday January 19, 2023 @09:09AM (#34440403)

> > Because the hive mind says AMD GPUs are better if you run Linux.

> As usual the hivemind just regurgitates memes without understanding.

I can't edit the OP any more, but I meant 3d drivers wise generally, not ML wise.

I haven't cared about 3d on Linux in ages, but now I do (thanks Apple for the architecture change) and need to decide what video card to get in the next year as prices go back to a reasonable-ish level. As in AMD or NVidia, not what price/performance tier. And my general impression from the internets(tm) is amd drivers are more stable. Am I wrong?

[ Parent | on HN ]

Re:GPUs for Deep Learning in 2023 – An In-depth Analysis by paol on Thursday January 19, 2023 @09:32AM (#34440654)

AMD drivers are not more stable. They are however open-source and included in the kernel, which has both advantages and disadvantages:

- AMD has better support for the "blessed" Linux 3D stack. This manifests itself as more mature support in Wayland compositors for example, which is still rockier with nvidia.

- With AMD you need to care about running recent kernels, because improvements to the drivers are tied to the kernel, as is support for newly released cards. With nvidia it doesn't matter - I can run a stable distro like Ubuntu 20.04 LTS and still use the latest driver.

- Nvidia support for switching between discrete and integrated graphics in laptops used to be crap. I'm not sure if it's still the case as I don't follow this closely, but given historical record, it's likely still crap.

- Nvidia driver quality is historically better than AMD (across the board, not just in Linux). If gaming is your primary use, things are still more likely to "just work" on nvidia than AMD, although AMD will usually fix problems eventually.

[ Parent | on HN ]

Re:GPUs for Deep Learning in 2023 – An In-depth Analysis by nottorp on Thursday January 19, 2023 @10:02AM (#34441013)

Hm so it's basically the same as the last time I cared. Thanks.

> Nvidia support for switching between discrete and integrated graphics in laptops used to be crap.

That's okay, I'm staying with Apple for laptops. I'm just setting up a Linux/x86 box in addition to the Apple desktop.

[ Parent | on HN ]

Re:GPUs for Deep Learning in 2023 – An In-depth Analysis by muxr on Thursday January 19, 2023 @08:49AM (#34440197)

Yes you can. And AMD's Linux driver is much better.

[ Parent | on HN ]