Hacker News Reader

Comment by pixelpoet | Hacker News Reader

AMD's rather old Radeon VII can really smash a 2080 Ti if you're doing relatively simple streaming computation, like say video processing. That 1TB/s HBM is no joke, and there's 16GB of it, at ~half the price of the 2080 Ti with 11GB.

Nvidia's memory compression is very impressive and allows them to do better with a narrower memory bus in many situations (such as games), however if your data is not so compressible (often the case in GPGPU) this advantage disappears.

Really all I care about is a good OpenCL driver implementation, especially OpenCL 2.2 support would be extremely attractive.

View on HN

> Really all I care about is a good OpenCL driver implementation, especially OpenCL 2.2 support would be extremely attractive.

nVidia seem to still only support OpenCL 1.1.

Presumably this is a conscious decision - they'd rather invest in CUDA.

I have no idea what AMD is up to regarding OpenCL. Have they just quietly dropped all support? [0][1]

Intel seem to be doing better, with support for OpenCL 2.1, with support for both CPU and GPU targets. [2]

[0] https://community.amd.com/thread/232600

[1] https://community.amd.com/thread/233197

[2] https://software.intel.com/en-us/articles/opencl-drivers

> nVidia seem to still only support OpenCL 1.1.

This hasn't been true for quite some time now. They support 1.2.

Interesting - their website says 1.1 [0] but they apparently released a driver in 2015 which supported 1.2 on certain chipsets [1].

[0] https://developer.nvidia.com/opencl

[1] https://stackoverflow.com/a/29607373/

The Radeon VII is also pretty fast in training transformer networks. In some models that I trained, it was only ~25% slower than our Quadro RTX 5000s.

On Tensorflow ROC (or whatever the name is?)

Yes, I used Tensorflow from AMD's ROCm repositories. However, the performance difference compared to upstream Tensorflow was not very large anymore.

I recently purchased a laptop with an Nvidia/Intel composite setup.

As a Linux user, I’ll never buy a laptop with an Nvidia GPU again (if I can help it). What a pain.

A nice piece of hardware with relatively pathetic software to drive it.

I hear that, I had a laptop some years ago with an Nvidia Optimus card in it. It was quite the frustrating experience. I got it to work well enough with bumblebee on Ubuntu but it never felt solid. I had hoped maybe the experience had gotten better but I guess not. I'll continue steering clear.

Nvidia has traditionally been very easy to use with Linux, as long as you were ok using their non-free drivers. Why was it so problematic for you?

> Nvidia has traditionally been very easy to use with Linux...

Unless you want to do some cutting edge stuff with it. We're begging nVidia to fix their DPMS implementation over DisplayPort for years now.

Oh, I forgot that nVidia drivers do not make enabling VSync easy for video/compositing overlays for ~6 years? I have to disable all optimizations and lose 50% performance to enable VSync in a compositing or video layer.

Yes, most of it works but some annoying parts do not work, and nVidia doesn't make anything easier.

This is on a GTX680 card. It was king of the hill when I first bought it.

Nvidia used to be the less bad option, but AMD's transition to AMDGPU leapfrogged that a long time ago.

They're still focused on their proprietary drivers, and naturally that means that they're way behind on supporting modern kernels. That's an issue if you have other hardware that only had support added recently. Their integration with X11's autoconfig also tends to have.. problems.

Nouveau basically doesn't work for their modern GPUs. My work laptop has a 1660Ti, and X11/Nouveau still can't render anything with it. Wayland/Gnome/Nouveau kinda sorta works, but I'd rather not have to use Wayland if I can avoid it.

Nvidia's support for PRIME is lacking. Their official stance was that you should restart your X11 session every time you wanted to swap. Bumblebee kind of worked, but was very slow and unstable. It looks like they finally implemented support for dynamic PRIME now, but of course they had to name their trigger environment variable something else, so that existing scripts for AMD PRIME wouldn't work.

In my experience Nvidia on Linux is fine… as long as you have a typical/common setup. The moment you step outside the norm, the experience falls apart rapidly and Intel/AMD become vastly more pleasant to work with.

I think using pop-os can solve the problem

Thanks for the suggestion, but I prefer NixOS

As one of the other replies to this comment mentioned, this is one of the pain points that the company System76 has spent some time working on. I think it's because it addresses a core part of their target market: professional developers or other users who might need access to the GPU for programming, but don't have the time or motivation to learn how to play around with a CUDA installation from scratch (I'm making the assumption that's probably a large part of your problem, as it has been mine).

Their Ubuntu-based distro, Pop!_OS, does a lot of good in this regard, and is really pretty plug and play for an Intel/NVIDIA combo set-up. The driver that they use to manage their CUDA install is also available for plain Ubuntu[1], and I've had good experience with using that to manage my own set-up running an Intel Xeon/NVIDIA P-series (if I'm remembering correctly). I haven't given it a shot on Debian, and if you're using something like Arch, there's a decent chance you'll be out of luck.

Note: I'm not affiliated with System76 in any way, just have slammed my head into this problem on multiple occasions and found this to be my go-to way of managing it.

[1] https://support.system76.com/articles/cuda/

Did you manage to achieve 1TB/s?

https://www.sisoftware.co.uk/2019/08/01/amd-vega2-gpgpu/ BW test over there shows 627GB/s of actual BW.

Not saying that it can hit 1TB/s, but I wouldn't put too much trust in one benchmark.

I got 704 GB/s on my machine[1] using their test.

[1] https://imgur.com/a/tq3nEz5

Reply on news.ycombinator.com