Hacker News Reader

In recent years, Linux computers have evolved to a major revenue source for Nvidia thanks to deep learning. However, not desktop users are behind this but servers, due to Nvidia's proprietary CUDA API. If they open sourced it or rebuilt it on top of mesa, it'd make it easier for AMD to implement CUDA, getting access to the deep learning ecosystem that's currently locked into CUDA. Nvidia's sales would take a huge drop. So I think it's even more likely that their drivers remain proprietary.

Reality check?

https://github.com/ROCm-Developer-Tools/HIP https://github.com/ROCmSoftwarePlatform/hipDNN

CUDA just works.

ROCm exists but many frameworks don't support it.

https://ai.stackexchange.com/a/16541

PyTorch supports ROCm quite well if you can follow three step build instructions instructions (https://lernapparat.de/pytorch-rocm/). TVM.ai also supports ROCm well. TF I didn't check.

FP32 performance for the VII is comparable to GTX20280Ti e.g. on Resnet training. Tensor Cores are cool, but 90% of people don't use FP16 that much.

I'm not sure that is the best reply that stackoverflow has ever seen.

The problem is that ROCm does not support many AMD compute cards, like their top of the line Navi-based one-year-old 5700 family.

Most frameworks do support ROCm though, so the ball is on AMDs court.

I can only speak for myself, but my understanding is that the Navi cards are not intended nor advertised for compute. Now, one might wish to get a consumer card with Arcturus after the Instinct MI100 is released or dislike that the Radeon VII is discontinued, or even that it wasn't split into render and compute lines, but basically, to me the complaint about Navi not being supported by ROCm sounds a bit like buying a boat and complaining that it can't fly. Driver support was bumpy for me with Linux 5.0 or so, but I haven't had trouble with just running whatever kernel Debian ships for quite a while now. But I must admit I don't use the GPU for anything but compute, so I don't know about graphics.

But your mileage may vary, I don't want to make people swear.

a GPU is a GPU, it's all just programs running on the shaders.

If AMD is unwilling to properly support its consumer GPUs in GPGPU workloads and NVIDIA is, then well... sounds like a good reason not to buy AMD.

It's really not a major ask, people have been running GPGPU programs on consumer cards since forever, even though they're "consumer" and sold as graphics cards and not compute accelerators. NVIDIA basically does this on purpose, people use their personal hardware to get a foot in the door of the ecosystem and end up writing programs that get run on big compute GPUs which NVIDIA makes big profits on, it's very intentional.

If AMD chooses not to do that, well... can't really blame people for avoiding their stuff when NVIDIA is willing to let you do this stuff on their cards and AMD isn't.

Call me back when ROCm supports their flagship graphics card.

The 5000 series has been out for an entire year without ROCm support at this point.

That is because AMD split graphics vs. compute and you look at a non-compute card. The Vega-based chips work quite well and the Radon VII is (or was, I'm afraid) an excellent value proposition.

So this is well outside my area of expertise; but that seems weird. I want AMD to shepherd the ecosystem to a point where I can run PyTorch with some support from a graphics card. Supporting some graphics cards and not others doesn't sound very promising.

It is amazing to watch how much of a struggle AMD is having with getting PyTorch to work with ROCm. It makes me appreciate what a good job Nvidia must have done with CUDA.

Nvidia has split lines too. But they understand the importance of getting developers on their platform.

(To be fair AMD has it planned, but it's behind several other priorities to them.)

> Vega-based chips work quite well

LOL what a fucking joke. Not even close.

Also that one VFIO bug makes one of the great advantages of Radeon disappear and there is no fix in sight.

Fuck AMD. Bunch of marketing hype and they ship garbage half baked crap to the market.

Yeah... I learned this the hard way.

I thought AMD had great Linux driver support, and open source drivers.

Turns out AMD only has the open source driver part of the things, and this driver does not support a large chunk of their products.

I'd take a binary driver over no driver any day (and in fact, I returned the amd card and am using intel now, which works great).

difference between AMD and nvidia:

AMD: some open source drivers, but many things are inop

Nvidia: all closed source, but everything works

Source: I have a Vega 56 and I am gonna give it the sledgehammer when 3000 series arrives. Fuck that shit to high heavens my dude. ROCm is buggy as fuck. Only the bare basics work on Linux. Drivers are buggy, and crash all the time. Even the most basic bullshit is not implemented right. Like there is no fan control. That's how bad it is.

I don't have so much of a problem with CUDA staying closed, but rather Nvidia sabotaging Nouveau through signed firmware which they don't release (and obfuscate in their blob). Nouveau would be probably be decent by now (not as fast or feature complete, but usable for real workloads on newer cards) if it weren't for the fact that Nvidia has added features which have the direct effect of making it impossible to have a competitive open source driver.

Maybe something will change on this soon. There was speculation about this: https://www.phoronix.com/scan.php?page=news_item&px=NVIDIA-O.... But I'm not holding my breath, and it would be nice if the solution wasn't "wait and hope until Nvidia releases the software necessary to control their GPUs".

> I don't have so much of a problem with CUDA staying closed, but rather Nvidia sabotaging Nouveau through signed firmware which they don't release (and obfuscate in their blob)

Do you have more info on this ? There is a big difference between not supporting open source, and actively sabotaging it. What are they doing, exactly ?

Look up "nouveau signed firmware". Phoronix has a bunch of articles on it. The Nouveau developers also talk about it at FOSDEM 2018 (and probably a later conference). This comment is a good intro: https://www.phoronix.com/forums/forum/linux-graphics-x-org-d...

TL;DR starting with the 9xx series, Nvidia started making it so their GPUs would only run firmwares signed by them (likely to prevent counterfeits, i.e. 2060s sold as 2080s). So it is impossible to control the fans and reclock the GPU. There's no workaround, so even as the person who owns the device, I can't run my own firmware. AMD has signed firmwares too, but they actually release sufficient blobs to fully run the device.

Wow, that's insane, thanks for sharing this.

When I had nvidia hardware, I never tried nouveau, so I never ran into this.

Yes, they only can do only one good thing - go bankrupt. That would be the best news for open-source.

Should open-source perhaps drop NVidia, instead of the other way around?

This divestiture happened long ago, if your running Linux not on a server, then Intel and AMD are the only players making chips that work right out of the box.

Even there though, only Intel has a buttery smooth experience. Ryzen for laptops is half baked (terrible USB C docking performance, occasional crashing on Windows and Linux with the 2xxxU series CPU/GPU chips) and AMD GPUs still require manual intervention to load proprietary firmware.

AMD does make some performant mobile GPUs though, they work well in Debian!

Funny, I stick with NVidia because they still offer more reliable, performant drivers for Linux (or these days FreeBSD), open source or no.

How so? Aren't you still trapped on Xorg with their proprietary drivers?

I switched to Wayland a few years back as vsync is quite nice to have, but whenever I go back to Xorg for AnyDesk or TeamViewer on AMD or Intel there is a fair bit of tearing.

Nvidia could have a competitive open source driver tomorrow if they released redistributable firmware that allowed reclocking the GPU.

> How so? Aren't you still trapped on Xorg with their proprietary drivers?

As a FreeBSD user I'm still "trapped" on Xorg anyway, and in any case I'd rather stick with what works than learn some whole new way of doing things for marginal benefits.

This is the sort of mentality I left the *BSD world for. Xorg is not some elegant codebase that's being unfairly targeted by an upstart.

> Xorg is not some elegant codebase that's being unfairly targeted by an upstart.

I never said it was. I said it's a whole new way of doing things for marginal benefits; is that not accurate?

You're not. I've been using Wayland exclusively on FreeBSD for a couple years now, with both AMD and Intel GPUs.

How do you go back to Xorg? Logout/login and/or reboot or is there a better way? I'm asking this because I often need screen sharing to work with my customers (demoes, check problems, etc) and I'm sticking with Xorg because Wayland doesn't do screen sharing.

Or even if they simply stopped signing their firmware. Nouveau has been perfectly capable of writing their own firmware.

Intel's drivers are far more reliable in my experience.

Slightly more reliable, but they don't make any cards that are good enough for gaming with IME.

They might work, for some definition of working, now getting the full capabilities of the GPU is another matter.

OpenGL 3.3 when the proprietary drivers do OpenGL 4.1 for example.

https://www.gpuzoo.com/GPU-AMD/Radeon_HD_6320_IGP.html

> then Intel and AMD are the only players making chips that work right out of the box.

I have an Intel GFX card. Had an AMD 5700. Linux support sucked, switched back to intel.

So AFAICT, only intel and nvidia have good Linux driver support, and only intel has good open source linux driver support.

You don't even have to go to open source. You can see this hostile behaviour from their top-paying clients!

Microsoft own previous gen Xbox emulator on the next gen xbox (i think it was original xbox emulated in the 360, but i might be wrong) was impacted by the team having to reverse-engineer the GPU because nvidia refused to let the emulator people to have access to the documentation provided to the original team.

Stumbling around Google didn't find me much more info on this, do you have any citations or keywords I could follow up on?

Which is funny because AMD drivers on linux has been nothing short of trouble (both open source and proprietary) and Nvidia’s blob just works.

[citation needed] My experience directly conflicts with this and IIUC most GNU/Linux users have exactly the opposite impression. Maybe you're thinking of some past situation?

No I am thinking of the situation where the open source driver doesn’t have opencl support and the AMD drivers (fglrx) doesn’t compile or requires dependencies so old that it was dropped by the packagers for Arch Linux, all this for a few years until they come out with something that actually works when I’ve never ever had an issue with Nvidia. Also AMD never ever figured out how to fix screen tearing, or do so in a sane that doesn’t involve trial and error editing xorg.conf.

Even on Windows AMD drivers are the most unstable, bugged software that’s even been shipped. It’s been a long standing joke that AMD “has no drivers”.

Yeah, fglrx driver with the issues you mention, offered OpenGL 4.1 support on a graphic card I have.

The open source driver replacement only does OpenGL 3.3.

I guess I should be happy it does any OpenGL at all.

Here, my card on the travel netbook that I use,

https://www.gpuzoo.com/GPU-AMD/Radeon_HD_6320_IGP.html

The open source driver is kind of ok if the only thing we expect from it is getting a working X session.

Now if one wants to do some complex OpenGL stuff, then it might work, or not.

That card is super old, before GCN. Have you tried with something a little bit more recent?

I used to have a HD 7950 and it always worked perfectly, same with my current Vega 56.

Sure, how do I replace the card on an otherwise perfectly working laptop?

The usual linux answer to hardware problems, keeps being to buy new hardware.

I would say that's a much more prevalent attitude in the Windows and Mac worlds. Linux tries to keep compatibility with really old software. It was only 4 years ago that major distros started to require at least a 686, aka Pentium Pro, released in November 1995!

But at some point you have to consider if it's really worth it keeping a 10 year old laptop around. It's painful to say them goodbye, I know, I have been there, but for me it's just not worth it.

Asus sold the laptop with Windows 7 support as well, the drivers kept being updated up to Windows 8.1, and thanks to Windows driver ABI, those drivers work perfectly fine in Windows 10.

No need to throw a perfectly working laptop to enjoy the DirectX 11 and OpenGL 4.1 capabilities that it was sold for.

Take any AMD 5000 series card. I had a top of the line 5700.

Still no driver for compute 1 year later. I'm so happy i decided to return it and switch to intel instead of waiting for AMD or some random joe on their free time to add support for it to their open source driver.

So yeah. I'd take a working proprietary driver over no driver any day.

Sadly I have to agree. Ryzen 3400G here, getting hardware transcoding on the iGPU is something I still haven't sorted out. There have been several recent issues in kernel, AGESA firmware (I suspect there might be newer versions with potential fixes that my mobo manufacturer hasn't released yet; this is 1.0.0.4 Patch B) and drivers. I've had several rounds of hunting down and compiling various versions of driver packages, modules and kernels from source, trying third-party PPAs, to no avail. The amdgpu/amdgpupro mess adds another layer of confusion.

I am not sure if I am missing some update, need to set some undocumented kernel flag and/or BIOS setting, if it's a software issue or Í just made a mistake somewhere. Debian 10/11.

Meanwhile, as much as I wanted to get away from Intel, their drivers have never posed any issue at all.

I believe AGESA 1.0.0.4 Patch B broke something for the APUs, you should try either upgrading or downgrading your BIOS but what worked for me was downgrading to AGESA 1.0.0.3 ABB, both Windows and Linux has stopped crashing now, although I still get the occasional lockup when browsing with Firefox on Linux. I found out the culprit after stumbling into this thread: https://old.reddit.com/r/AMDHelp/comments/gj9kpz/bsod_new_pc...

It's anecdotal of course, but my RX560 has been absolutely flawless on both Ubuntu and openSUSE, literally out of the box support on a standard install.

Ryzen 3700X and Radeon 5700 XT here. Not a single problem.

I have recently used Radeon 550, 560, 570, 5500 on AMD 5050e (yes, that old!), Ryzen 1600 (non af), 3100, 3600 and all have worked fine, Ubuntu 16.04, 18.04 and 20.04. In fact on average I have found the various hardware configurations to be about 5% faster on Linux than Windows.

> Quoting Linus Torvalds (2012) [0]:

Is this an Ad Hominem ? Linus does not mention there a single thing that they are actually doing wrong.

> Drew DeVault, on Nvidia support in Sway [2]:

Nvidia has added wayland support to both KDE and GNOME. Drew just does not want to support the nvidia-wy in wl-roots, which is a super super niche WM toolkit whose "major" user is sway, another super super niche WM.

Drew is angry for two reasons. First, sway users complain to them that sway does not work with nvidia hardware, which as a user of a WM is a rightful thing to complain about. Second, Drew does not want to support the nvidia-way, and it is angry and nvidia because they do not support the way that wl-roots has chosen.

It is 100% ok for Drew to say that they don't want to maintain 2 code-paths, and wl-roots and sway do not support nvidia. It is also 100% ok for nvidia to consider wl-roots to niche to be worth the effort.

What's IMO not ok is for Drew to feel entitled about getting nvidia to support wl-roots. Nvidia does not owe wl-roots anything.

---

IMO when it comes to drivers and open-source, a lot of the anger and conflict seems to steem from a sentiment of entitlement.

I read online comments _every day_ of people that have bought some hardware that's advertised as "does not support Linux" (or Macos, or whatever) being angry at the hardware manufacturer (why doesn't your hardware support the platform that says it does not support? I'm entitled to support!!!), the dozens of volunteers that reverse engineer and develop open source drivers for free (why doesn't the open source driver that you develop in your free time work correctly? I'm entitled to you working for free for me so that I can watch netflix!), etc. etc. etc.

The truth of the matter is, that for people using nvidia hardware on linux for Machine Learning, CAD, rendering, visualization, games, etc. their hardware works just fine if you use the only driver that they support on the platforms they say they support.

The only complaints I hear is people buying nvidia to do something that they know is not supported and then lashing out at everybody else due to entitlement.

What exactly are you invested in in this discussion? We were originally discussing Nvidia's business practices don't match ARM's business practices. But you seem to just want to take on people's personal views on Nvidia now.

You're now somehow arguing with people that they should stop complaining about Nvidia's business practices. I would agree with that in the sense that Nvidia can do whatever they want: nobody is obliged to buy Nvidia, and Nvidia is not obliged to cater to everyone's needs. It's a free enough market. But even if you don't agree with some/most of the complaints surely you must agree that Nvidia's track record of pissing of both other companies (and people) is problematic for when they take control of a company with an ecosystem driven business model like ARM's?

> I would agree with that in the sense that Nvidia can do whatever they want: nobody is obliged to buy Nvidia, and Nvidia is not obliged to cater to everyone's needs. It's a free enough market.

I'd agree with you this is OP's argument, however it's main flaw is in explicitly omitting the fact that NVidia is not the only party that's "free" to do things.

We're not obliged to buy their cards and we aren't obliged to stay silent regarding its treatment of the open-source community and why we think it would be bad for them to acquire ARM.

I am always amazed at the amount of pro-corporate spin from (presumably) regular people who are little more than occasional customers.

> We were originally discussing Nvidia's business practices don't match ARM's business practices.

We still are. I asked about "which specific business practices are these", and was only pointed out to ad hominems, entitlement, and one sided arguments.

Feel free to continue discussing that on the different parent thread. I'm interested on multiple views on this.

> You're now somehow arguing with people that they should stop complaining about Nvidia's business practices

No. I couldn't care less about nvidia, but when somebody acts like an entitled choosing beggar, I point that out. And there is a lot of entitlement in the arguments that people are making about why nvidia is bad at working with others.

Nvidia has some of the best drivers for Linux there are. This driver is not open source and distributed as a binary blob. Nvidia is very clear that this is the only driver that they support on Linux, and if you are not fine with that, they are fine with you not buying their products. This driver supports all of their products very well (as opposed to AMD's, for example), its development is made by people being paid full time to do it (as opposed to most of their competitors which also have people helping on their drivers on their free time - this is not necessarily bad, but it is what it is), and some of their developments are contributed back to open source, for free.

People are angry about this. Why? The only thing that comes to mind is entitlement. Somebody wants to use an nvidia card on Linux without using their proprietary driver. They know this is not supported. Yet they buy the card anyways, and then they complain. They do not only complain about nvidia. They also complain about, e.g., nouveau being bad, the Linux kernel being bad, and many other things. As if nvidia, or as if the people working on nouveau or the Linux kernel for free on their free time owes them anything.

I respect people not wanting to use closed source software. Don't use windows, don't use macosx, use alternatives. Want to use linux? don't use nvidia if you don't want to.

No, thank you. Can we do without the Reddit attitude here please? The choosing beggars, the needless quoting of debate fallacies in what is not a debate?

If you ask your impression of Nvidia's business practices, and they give you their opinion, you can't somehow invalidate that opinion by retorting with debate fallacies. That's the "fallacy fallacy" if you're sensitive to that. This is not a debate competition about who's right, this is people giving their opinions based on Nvidia's past and current actions. You asked a question, and they answered. This is not a competition. Please give them the basic respect of acknowledging their opinion.

This is so wrong.

We are discussing a topic, and people throwed multiple arguments that do not make sense.

You are claiming that I should just shut up and respect their feelings, but that is worthless.

Two examples:

---

Somebody's argument was: "Linus doesn't like them, therefore I don't like them".

The reason these are called logical fallacies is because these arguments are illogical. I told them that this was a logical fallacy (argument of authority - just because someone with authority makes an argument does not mean they are right), and ask them _why_, what is it that linus and you do not like.

I am happy I did that, because many of them have raised multiple actually-valuable arguments in response. For example, because nvidia's hardware throttles down if the driver firmware is not signed, and this makes the open source drivers slower for no reason.

That's a valid and valuable argument. Linus doesn't like them is worthless.

The person who raised this argument learned something from somebody else which knew what Linus did not like, and so did I.

---

The same happened when I called out the entitled choosing beggars. "Why are you angry at nvidia for not providing an open source driver ? You knew before buying their product that only the binary driver was supported."

Read the responses. The reason they are angry, is because they don't have a choice but to use nvidia, because the competition products (AMD in those cases) are much worse. AMD does have open source drivers, but they are crap, and they don't support many of AMD's products, at least for compute, which is something that many (including myself) use for work.

These people have picked a platform that values open source code, but due to their job requiring them to actually get some work done, they must use nvidia for that, and they don't like having to compromise on a proprietary driver.

Honestly, I think this is still entitlement, but I definitely sympathize with the frustration of having to make compromises one does not like.

---

From the point of view of whether nvidia buying ARM is good or bad. I still have no idea. ARM does _a lot_ of open source work, its major market are Android and Linux communities.

I understand that people are afraid that Nvidia will turn ARM into a bad open source player. It can happen. But without Android, iOS, and Linux, ARM is worthless. So a different outcome of this could be that NVIDIA buying ARM ends up making NVIDIA more open source friendly, since at least the Linux market is important for nvidia as well (~50% of their revenue).

It definitely makes sense for regulatory authorities to only allow somebody to buy ARM that will preserve ARM deals with current vendors (apple, google, samsung, etc.), and that also will preserve ARM open platform efforts.

If nvidia does not agree to that, they should not be allowed to buy arm.

> their hardware works just fine if you use the only driver that they support on the platforms they say they support.

Except that they stop supporting older HW at some point. That, together with occasional crashes learned me not to buy nVidia HW again.

> Nvidia has added wayland support to both KDE and GNOME.

NVIDIA insisted on pushing its own EGL streams even as the wider community was moving in a different direction.

They suffer from a major NIH syndrome and do not know how to work with others at all.