AMD Linux Graphics No Longer Unusable For Blender Developers: 251 To 9 Seconds Speed-Up

[qarium]

Obviously its just substandard in every respect.
Can't do blender

i tested it 30 minutes ago and Blender version 3.5.1 works in HIP mode detecting 2 devices my cpu and my vega64 gpu
and i am on Fedora 38 and only used the command dnf install rocm*

this means AMD ROCm/HIP works for Blender3D 3.5.1 out of the box on Fedora38 with just installing the rocm packages.
my vega64 is from 2017 this means its a more than 6 year old card.

Can't do games (e.g. GTA V)

I support people who can not play their favored games on windows in one case sims2.
i installed her Fedora38 with Lutris and sims2 did work out of the box with the lutris game loader.

if you search for GTA V in the Lutris database result is this

https://lutris.net/games/grand-theft-auto-v/

of course it works...

a $1000 AMD GPU is on par with a $200 Nvidia GPU for compute, e.g. the jetson nano

The most important single factor what a GPU can do or not is the "Memory Wall"
a 1000 dollar AMD gpu has 24GB vram of course this is better than a jetson nano with 4GB ram...

https://www.youtube.com/watch?v=GuV-HyslPxk

this video only has historical information and say nothing about the situation right now or the future.

And while yes, the 7900XTX (and only the 7900XTX) is finally a reasonably "competitive" GPU from AMD, after something like a decade or two of them not even trying.

for AI/machine learning because of the Memory Wall there are multible much better AMD gpu's/compute accelerators from AMD.

for example AMD Instinct Mi100 it has 32GB vram HBM2 with 1228,8 GB/s and 7690 shader cores

https://www.mindfactory.de/product_info.php/32GB-AMD-RADEON-INSTINCT-MI100_1413633.htm

even to the fact that this card is much older and only 7nm compared to the 6nm of the 7900XTX it easily outperforms it in any compute workloads. there are 2 reasons for this the 7900XTX has only 6144 shader cores and only 960GB/s
as you can see if you want to spend more money and you have only compute workloads you get a faster card with more VRAM from AMD.
this card outperforms the 7900XTX in Blender with HIP easily.

now you say i do not only have compute workloads i also have 3D workloads well for AI/machine lerning stuff you still have a better card as the 7900XTX from AMD... the AMD Radeon Pro W7900 it has 48GB vram and this means memory wall is less of a problem.

with that 48GB vram you can run very large AI modells.

If you don't need what nvidia is selling for your desktop
You'd be better just picking up an RTX3050 for 1/3 of the price, having access to what nvidia is selling, for no noticeable performance difference from that 7900XTX and none of the headaches that come with AMDs low effort desktop GPU drivers.

the low effort desktop GPU AMD driver run's out of the box on fedora 38 on my 6 year old Vega64 with Blender and other apps and of course also with pytorch HIP backend.

why should i buy a 3050 if my vega64 already does everything i want ?

And if you don't care about the headaches, just grab a A770 from intel for $350
And if all you want to talk about is games, just buy a games console and STFU.

can you tell me what headaches you talk about ? ROCm/HIP works in Fedora38 out of the box with simple type in dnf install rocm*

[mSparks]

of course it works...

Uhuh....

What exactly do you get for giving AMD that $1000 again?

can you tell me what headaches you talk about ? ROCm/HIP works

In the same way GTA V "works" you mean?
Or do you mean, it will work, eventually, probably, just trust you, it will be the new standard eventually, pinky promise.
Like wayland...

Wake me in another 20 years, maybe you'll be right, and AMD will be worth buying in 2044.
Until then imma stick with nvidia, that I know and see substantial investment in the software that drives their hardware.
Rather than just throwing any old garbage to the community and letting them solve all the hard problems on their own.

[qarium]

Uhuh....

What exactly do you get for giving AMD that $1000 again?

I don't think such a performance erratum is there for long time.

In the same way GTA V "works" you mean?
Or do you mean, it will work, eventually, probably, just trust you, it will be the new standard eventually, pinky promise.
Like wayland...
Wake me in another 20 years, maybe you'll be right, and AMD will be worth buying in 2044.
Until then imma stick with nvidia, that I know and see substantial investment in the software that drives their hardware.
Rather than just throwing any old garbage to the community and letting them solve all the hard problems on their own.

wayland is already a standard ...

you can do whatever you want... but some people here in the forum report that they now use the integrated GPU of the mainboard/CPU to get the AMDGPU driver for the desktop and they only use a Nvidia GPU for CUDA ... this solution they say works best.

stradegies like that show that people exit the Nvidia ecosystem and isolate Nvidia to the absolut minimum like CUDA..

[smitty3268]

In the same way GTA V "works" you mean?
Or do you mean, it will work, eventually, probably, just trust you, it will be the new standard eventually, pinky promise.

Pretty sure it works fine on AMD's drivers. This is showing a bug in radv only.

See for example: https://www.phoronix.com/review/amdvlk-radv-rx7900/4

Numbers aren't comparable because the setup is different, but radv is getting smoked in that test.

It's obviously not ideal that such a bug exists, but on the other hand if that is the one app you really care about in terms of performance it's not that difficult to run the alternate driver for it.

[Magissia]

Nope, it's not evil, there is nothing wrong with capitalism.
But encouraging people to do stupid things (like pay extra for substandard desktop hardware by AMD) because AMD are somehow better for releasing meaningless bits of their GPU drivers first as GPL is pretty evil.

Point of view.

I prefer open source gpl drivers because they can be included in the kernel directly which makes my life easier. To each their use for that.

[mSparks]

Point of view.

I prefer open source gpl drivers because they can be included in the kernel directly which makes my life easier. To each their use for that.

Long ago a solved problem

https://rpmfind.net/linux/rpm2html/search.php?query=akmod-nvidia

https://src.fedoraproject.org/rpms/akmods

[shinger]

now as Nvidia user with KDE:
- AD1. flickering - reason is entirly diffrent on X and Wayland. X it is simply allowing tearing by default and some settings like triple buffering to full composistion pipiline removes it, or simply using gsync/freesync etc. On Wayland well it is really complicated situation mostly due to explicit vs implicit synchronization model. But nvidia (and not only nvidia) works on allowing Wayland to support explicit synchronization present to solve it,
- AD2. I agree with KDE updates, common issue, heck even restart button in menu doesn't work after update, you need to write reboot in terminal and after reboot stuff works again, but yea really common hassle,
- AD3. You can turn that question other way around - what if you use some distro with older kernel (idk let's say Ubuntu 22.04 LTS very popular distro choice) and it doesn't have support for newest AMD GPU out of box. Well in case of Nvidia you just install newest driver (and Nvidia supports newest hardware pretty much day 1 on linux mostly) without forcing you to pick some extremly fresh kernel, so I agree there are issues with Nvidia way of updating, there are also benefits to it,
- AD4. I didn't expierience that on Kubuntu.
- AD5. Didnt' expierience that,
- AD6. Never used Viber and never installed such stuff through Flatpak. But discord screen sharing worked for me.
That being said I used mostly Kubuntu default (X server on Nvidia). Tried Wayland a bit and it is kinda functional but with more issues for sure.

A lot of however Nvidia issues with Wayland (or linux in general) kinda originate from that open source people pick some design choices without much thinking can it really work right way for users. A lot of times it is "art for sake of art" not for sake of usability. Examples could be ignoring that world moved to explicit synchronization model while Linux/Wayland is left implicit, DMA-BUF marked as GPL-only symbol (making it more pain for Nvidia users again), ignoring Nvidia criticism for GBM and entire situation with EGLstreams (both solutions had own issues, but clearly with better cooperation we could have something better then both), and some issues are literally nature of OpenGL being horrible (like I remember bugs in KDE of sort hey after some operation state is undefined, MESA assumed one, Nvidia assumed another - of course KDE devs assumed MESA version, and Erik from Nvidia made PR to force set state to avoid that undefined behaviour.

Now Nvidia is most apparent to suffer from those issues. But not just Nvidia, I remember AMD and Intel developers complaining about implicit synchronization heck even Joshua from DXVK/Valve wants it, AMD propertiary GPU drivers probably don't enjoy some GPL-only symbols (yay double effort to maintain drivers!), I kinda remembered some buffer issue when AMD dev admitted they would need to use something similar to EGLstreams to solve something, OpenGL undefined behaviours are subject river... Things are getting better in general, but a lot of those issues could be avoided.

And regarding main subject in thread Nvidia would fix you those bugs, as they do. Explicit synchronization for present is done by Nvidia employe in Wayland/Xwayland. Erik had plenty of contributions to KDE as example and a lot of them were bug fixes. A lot of issues Nvidia-only comes to that no one tests dev like build on nvidia before shipping. So yes you do face some bugs, but I would say Nvidia seems to be doing pretty good considering no one tests new Kernels against Nvidia driver or new versions of Wayland/X server.

On my work laptop we use Linux as our OS (which i am happy with, but no happy with Nvidia chip). In the past i had Ubuntu, because it was much more stable. Last year i bought a 49 inch curved screen and to be able to work productively as being able to easily put different windows next to each other i started using KDE as KDE is the way to go forward when it comes to amount of productive features to configure on your desktop. Yes i did use Kubuntu, but Kubuntu had an issue with Forticlient that we use as a VPN client. Forticlient needed a dependency for it to work properly, but Kubuntu refused that dependency, but it needed a different version of that dependency. So the Desktop Interface says it wants version X and Forticlient wanted verion Y, after tinkering with it for 5 hours i came to know that. The problem was that Forticlient could be started, the logs were not verbose enough to find the real cause. Again a example of a rubbish application). Beside this issue i had enough drama with APT in the past that when i installed some application and one of the dependency of that application might be Ubuntu-Desktop and you wanted to remove that application, suddenly Ubuntu-Desktop also needed to be deleted. In the past i made this mistake.

I also thought about the future and Canonical forcing people to use SNAP. I mean lets take flatpak as an example. To use flatpak for lets say 2 or 3 applications i get it. But to use flatpak for all your applications i would not really want that. So i see them forcing it on you. Also talked about that the amount of depth of partnership that Canonical is having right now with Microsoft has made me to be careful and don't put all of your eggs in one basket. That also being said, that even today you have to work with outdated Linux kernels (Older LTS versions) for the sake of 1 company that refuses to work with the community, they are just asking consumers to not buy their products.

And no it is not like i am using Arch if you talk about "new", i am using the latest stable releases of the applications. Be it KDE Plasma, be it Linux kernel or other applications through repo's or Flatpaks.

You guys really cannot expect people to still be respecting a company while the company is not respecting you as a user or consumer. I mean oke..i am a nobody, but if Linus himself even is complaining about Nvidia that maybe there really is something going on. And this was even almost 6 years go..still a lot of issues with Nvidia.