Hacker News Reader

Comment by godelski | Hacker News Reader

I do graphics, but on the scientific side, we all use Linux. Except when the lab provides us Macs, then we get frustrated and just remote in.

I'm just trying to detangle the type of graphics being mentioned. I do wish Nvidia drivers were better, but maybe that'll happen with ROCs coming out soon.

View on HN

> I do wish Nvidia drivers were better

Get AMD. Why do people insist on getting Nvidia cards and then complain that those aren't aimed at them? Get the one that is selling to you.

Cannot reply to sibling... Nvidia doesn't have simply the best OpenGL implementation, it has the fastest and most forgiving implementation. No argument about speed (or Cuda), but by being forgiving of code that shouldn't work according to spec, it creates lock-in and makes more conformant implementations look bad.

Also, on Linux, AMD's OpenGL implementation - which is just Mesa - is very high quality.

> but by being forgiving of code that shouldn't work according to spec

So they are not following spec, and that's a good thing?

If you're trying to run a non-conformat application that you are unable or unwilling to modify, then yes. It is a good thing.

Because you simply can't do much ML experimentation on AMD; it lacks the entire CUDA ecosystem. I'm sorely tempted to run two GPUs, but also worried they'll probably clash.

I am interested to see ROCm. AMD has the contact for the exascale machines and there's a big push for ML at the DOE. So I hope there's some real competition coming (looks like it's already integrated into pytorch btw)

Because Nvidia has faster hardware (assuming you want that) and the best and most complete OpenGL implementation.

At least that is the reason i get Nvidia GPUs.

Except Nvidia is selling to me. I use cuda. At the end of the day a higher performing GPU is (MUCH) more important than running displays. There's lots of legacy code using cuda as well, so there's some lock in (AMD is tackling this issue but that's to be seen. We also use data parallel primitives - good for portability between super computers and things with different architectures - for many things but not everything, so when ROCm comes around this may not be an issue at all). But as of right now I'm locked in, at least with my priorities.

So please just stop telling us to get AMDs. That trades a small issue for a bigger one. Nvidia is selling is 90% of what we want and we're annoyed that they just don't care about that other 10%.

Because until recently, on Linux AMD's driver story hasn't been good and performance has lagged. The driver situation is improving but it will take time to win people back. It also doesn't help that they don't (yet) have a Ryzen-like story on GPUs.

Does the new driver architecture support controlling fan speed yet? I've been on both team green and team red, and not being a purist I have just used their proprietary drivers. It wasn't until AMD shifted to their new driver architecture (with most people suggesting to use the "high quality" open source one that required baking firmware into my kernel (I run Gentoo)) and I had so many issues with both the new open source one and proprietary one that I accelerated my time-to-update-card timeline and went back to NVidia, emerged nvidia-drivers, and have never had a problem since. Nvidia has been rock-solid for me on Linux (even when I first had an old card in 2009-2012 or so) using their proprietary drivers. There's also the continuing fact that while you can get 2 AMDs to match the performance (and cost) of one Nvidia, I'd rather just have the one card every time.

Don't know as I'm right there with you using nVidia myself right now. While I'm going to be picking up a new Ryzen-based system soon, AMD's GPU offerings aren't where they need to be for me to shift away from nVidia yet.

Because both times I bought an AMD (ATI at the time) card, it failed prematurely (out of warranty, but < 2 years of life) and the drivers sucked balls.

Where are you buying graphics cards with less than two years of warranty? If your GPU budget is that slim, you should probably be relying on integrated GPUs.

Probably in every place that is not the European Union which mandates 2 year warranties. Rest of the world iirc is 1-year.

I did check about five brands of AMD graphics card on Newegg US before posting, and didn't see anything shorter than two years, with three and five years quite common. I also once bought an AMD card with a lifetime warranty, though I believe that particular manufacturer has since stopped offering those terms on any of their models.

So the question remains valid: what are the GPUs being sold with super short warranties, and are they even worth buying in the first place?

The situation was much different 5 or 10 years ago.

One year was the best you'd get, 90 days was common.

Even on AMD you can have awful problems. The modeset xorg driver froze for me for a month until someone finally told me there's also the radeon xorg driver, awful.

May I ask what ROCs is and why is it making Nvidia drivers better?

Not ROCs, it's ROCm. AMD's GPU compute libraries. It won't make NVidia better, but will provide an alternative to those who are stuck with NVidia for doing tensorflow, PyTorch stuff.

https://rocm.github.io/

Creating scientific graphics is a much different job than doing graphic design.

Reply on news.ycombinator.com