Hacker News Reader

Comment by modeless | Hacker News Reader

Updating is going to be a huge pain. It always is, for one reason: Nvidia. I need CUDA for ML, and nouveau is so bad anyway that 18.10 wouldn't even boot on my machine when I tried to install it.

It's guaranteed that if I upgrade right now I will be troubleshooting black screens, hangs, and compiler incompatibilities for hours if not days in an attempt to install a working set of ML tools including up to date drivers and CUDA.

It's sad that one required proprietary component ruins the whole experience. But there's no other option for ML research. AMD just isn't investing enough.

View on HN

If you are using it as a working environment, you probably should stick to LTS releases.

In theory, things should be better[1] (I have an nvidia card too, I will probably hold off for a few weeks).

1: https://www.forbes.com/sites/jasonevangelho/2019/04/15/ubunt...

We've said "things should be better" since GLX became a thing. The real problem is that every Linux-based OS is too stupid to realize when something in the windowing system is broken. Every time I upgrade Ubuntu, I run into a login loop where Xinit refuses to work, gets sacked, then redirects me back to the login screen. The fix is a simple one liner. I see the solution to my problem has 11k upvotes. The solution is 5 years old.

It's been like this for almost a decade, and I haven't even gotten into Laptops with Integrated/Discrete "piggyback" setups. That nut still is not cracked.

It's kind of sad.

I have one of those laptops with integrated and discrete cards. Really, it works way better than you'd think; the official discrete GPU driver is a part of the kernel and gets better with each kernel release and requires no weird proprietary blobs, and running a program on the discrete instead of integrated GPU just requires running it with `DRI_PRIME=1` in the environment.

The main issues with GPUs in Linux rests squarely on nvidia's shoulders. Intel has worked flawlessly for ages, and after AMD actually started upstreaming their drivers, their cards have been a dream to work with.

Thing is, I need NVidia because of CUDA. It's so frustrating to be put in a situation where I have to deal with the sucky drivers because, where else am I going to go? For some silly reason, AMD still has a reputation of "sub-par" to some science circles, and to this day I have no idea why that is the case. I suppose it's hard to market "OpenCL", which is why Khronos is rolling it into Vulkan very soon. NVidia showed up with just as much marketing as they did a useful product.

And (please correct me if I'm wrong) OpenCL support in NVidia drivers is still in the "consideration" phases as far as I understand it.

Yeah, it's certainly an issue, because people do have nvidia cards and they don't work as well as they should have on Linux. I'm just pointing out that it's in large part nvidia's fault, not really Linux' graphics stack's. I myself have an nvidia card in my desktop, and while it's _mostly_ fine with the proprietary drivers, there are occasional graphical glitches which just don't happen with other cards. (The nouveau drivers just don't work at all, which is also largely nvidia's fault from what I understand, due to not providing many docs and forcing the nouveau folks to reverse engineer everything.)

Coincidentally, I decided to try getting into OpenCL a couple days ago. I've just used the Intel OpenCL runtime, but Arch Wiki's GPGPU page mentions opencl-nvidia: https://wiki.archlinux.org/index.php/GPGPU#NVIDIA

I don't understand why CUDA is that big, when it's locked down and supported by only one vendor, and OpenCL exists and is open and supported by all vendors. But then again, there are probably some market forces I don't understand; I also can't really grasp why Direct3D is that big when it's locked down and supported by only one OS vendor and OpenGL exists.

I upgraded my notoriously obnoxious NVIDIA GTX 970 with 4 GB VRAM and different RAM speed for one of the GB's with no hiccups here, and that was from 18.10 to the 19.04 beta. :) But of course, something always need to be said about backups before upgrading installs involving proprietary drivers. Next time I'm going either AMD or Intel for the GPU. This crap is on NVIDIA so as I'm planning full time Linux later this summer, I'm removing them from the equation.

Strange. For me, it was the installation and setting up of CUDA on my long-in-the-tooth Ubuntu 14.04 LTS box that led to me upgrading to 18.04 LTS (along with a hand-munged version of the latest gcc, plus an ill-advisde NVidia driver upgrade after both of those that hosed things).

That said, while I do have the latest NVidia drivers now (albeit on an old 750 ti/sc card), I haven't tried to re-install CUDA or other ML tools that I had before (it was done for Udacity's "Self-driving Car Engineer Nanodegree" - to compile code for the assignments, and to utilize my GPU to the fullest, I needed to do a bit of a mashup that subtly broke things - but it worked well enough for the course).

Maybe doing that might show something similar to what you describe?

I know when I did it the first time, I was kinda surprised at how easy it worked out; I had CUDA working with Tensorflow quickly, and it all played nice with the GPU rendering at the same time (we had to do things like train a model using data from a Unity-based 3D driving simulator, so when you ran the model to drive the car, it had to use both CUDA for the model and GL for the graphics, and amazingly it all worked on that poor 750 - and had a decent frame rate!)...

https://news.ycombinator.com/item?id=19282074

you could burn it to a USB drive, boot from that, and verify before you go crazy upgrading....

Reply on news.ycombinator.com