NVIDIA Confirms Plans To Drop "Kepler" GPU Driver Support

[Linuxxx]

You're right, OpenCL is the shame of AMD: https://github.com/RadeonOpenCompute...ment-860318228

The only remaining reason to turn away AMD GPUs today is compute. For someone not doing compute, there is no reason to buy Nvidia at all.
OpenCL with AMD is kinda working, but there is like, 3 complete drivers given your card generation (Orca, Pal, ROCm), and one incomplete meant to be universal (Clover/libCLC). In the end, it's like OpenCL does not exist because users will likely fail at installing them and/or making it works.

That's a shame because when we manage to get OpenCL working, it works, really. But the path to success is just so hostile…

We really need from AMD to complete and polish their OpenCL stack because we should not buy Nvidia, never.

Even if AMD would release stellar & bug-free OpenCL support as soon as tomorrow, it still wouldn't make much of a difference in the GPGPU space even in the medium-term, let alone short-term!

What you fail to realize is just how entrenched the CUDA way is in this market space...

Anyway, the only hope AMD has right now are the forthcoming supercomputers making use of their hardware & software; at least it is enough of a threat for nVidia to force them to play by Linux's rules and open-source & upstream their kernel GPU driver part like AMD has done.

[Aeder]

True, but the old driver doesn't disappear overnight, and is sometimes still updated for newer kernels. Case in point, Fermi cards are still supported on the 390 line of Nvidia proprietary drivers, and the Arch maintainers have tested it with the 5.12 kernel. https://wiki.archlinux.org/title/NVIDIA#Installation.

I'd still make the case AMD & Intel are better positioned to support old hardware. I've switched to AMD myself, and I do enjoy the seamless install/update experience not playing dkms-roulette anymore.

Not always though. My Intel+AMD hybrid graphics laptop went from booting fine to a login screen freeze with kernel 5.11 unless I run it with workarounds like radeon.runpm=0

[Paradigm Shifter]

Even if AMD would release stellar & bug-free OpenCL support as soon as tomorrow, it still wouldn't make much of a difference in the GPGPU space even in the medium-term, let alone short-term!

What you fail to realize is just how entrenched the CUDA way is in this market space...

Anyway, the only hope AMD has right now are the forthcoming supercomputers making use of their hardware & software; at least it is enough of a threat for nVidia to force them to play by Linux's rules and open-source & upstream their kernel GPU driver part like AMD has done.

Things can change quickly, though, given appropriate incentive. The rapid rise of CUDA never surprised me, as it was the only game in town at the time. But the rapid rise and domination of Facebook, for example, did surprise me, as there were already well established, popular competitors. Get the AMD compute stack into decent shape, do a couple of deals on new supercomputers to get Radeon cards in rather than nVidia... that'll get scientists interested in working on them more. I know a fair few people who aren't keen on being totally dependent on CUDA, so if alternatives came along, they'd be tested. Don't get me wrong, AMD is at a serious disadvantage against CUDA, but with some stubbornness and care, they could surprise us all.

If AMD can keep their dominance on the console market, and perhaps do some bundle deals to get a couple of new super computers loaded out with Epyc and Radeon, rather than Epyc and nVidia... it could get interesting.

Of course, nVidia isn't standing still either. Now they've purchased ARM, if what I predicted when the acquisition was announced - that nVidia go "all in-house" for a compute ecosystem; total control of CPU, GPU and OS/software (they were hiring some BSD devs a while ago, I think?) - happens, then that would go one of two ways, depending on how willing users are to be completely dependent like that... and how much nVidia increase prices if they do.

[mdedetrich]

This is why, guys, you should not buy Nvidia. With AMD you can run cards from 2002 (R300) with latest software… and it works.

With Nvidia, if you try to run cards from 2006 (Curie), you have to run Xenial (from 2016, which just leaved standard LTS) and because that Nvidia driver is buggy as hell, not use GNOME Shell. Nouveau not being usable on cards of this generation for other reasons.

Given how the game is built, a game may not run on 340 driver today, that means the Tesla cards. There was still new Tesla GPUs being released in 2011, yeah, while Fermi was already a thing…

And well, you'll say that's old? Yes, and now we talk about Kepler. My Laptop has a Kepler, I have no feeling I should update the laptop to something else, it's fine! I bought it second hand so Nvidia did not got my money, but eh, Nvidia is bad.



No, this is a lie to persuade oneself Nvidia is not that wrong. It's false, newer software also bring fixes for old hardware, and support usually gets better with time.

Have you ever done this in your life, running an aging distro to run previous generation hardware ? Have you lived with it ? The only reason to do so is to do retro-computing, but Kepler is not retro-computing.

You can still run latest Ubuntu with latest Firefox on early AMD64 and GPU from 2002. At this point it's almost retro-computing, I'm not saying we should but it's a fact, it works.

And then, on Nvidia side, I would have to run an aging LTS to use my current laptop with 8 threads CPU and 16 GB of RAM, and two SSDs ?

My laptop with Kepler (Quadro K1100M) is very fine, I don't think I have to stick on an aging LTS because of Nvidia. I have no reason to buy a new laptop.

A computer with a Kepler Quadro, 8 threads, SSD and 16 GB of ram is not retro-computing.

My Thinkpad W541 is really OK for today's use. Luckily it also has an Intel GPU so I can just forget about Nvidia deprecating my hardware… Fun fact, the Intel GPU is now faster at doing OpenCL than the Nvidia one… Nvidia seems to be designed to be sold, not really to be used.

On the other hand, AMD is still improving the GCN 1 support in amgpu/radeonsi today.

AMD GCN 1 GPUs were out in the same years as Kepler. GCN 1 have OpenGL 4.5, Vulkan, not retro-computing. And Nvidia is just dropping Kepler, their competitor to that generation ?

Actually to be more correct and to qualify, this is why if you care about old hardware you shouldn't buy NVidia + Linux. You don't have this problem on Windows, you can run 10 year old systems with the latest support NVidia binaries without any issues (I have such systems).

The core problem is that because Linux insists that drivers have to be in tree (rather than having a stable ABI like Windows does) means that when Nvidia releases a blob from some graphics cards, it typically only works for that generation of Linux kernels (typically its a range of one year). What ends up happening with 10 year old cards is that the Kernel has changed so much in that span of time that the DKMS wrapper module which NVidia uses likely needs to be completely rebuilt for the old driver.

The is the downsides of having a monolithic rather than hybrid (Windows/Mac) kernel, longtime driver support only works well if you its in the kernel otherwise you are basically f'ed.

Its not like NVidia doesn't care about old cards, they do (there are plenty of ancient systems running Windows that use NVidia hardware). Its just that its not practically feasible for them to do this on Linux because of what I stated earlier. With Windows this is a non issue.

[illwieckz]

Actually to be more correct and to qualify, this is why if you care about old hardware you shouldn't buy NVidia + Linux. You don't have this problem on Windows, you can run 10 year old systems with the latest support NVidia binaries without any issues (I have such systems).

That's right but on the other hand if you're ready to use or have to use Windows the Nvidia closeness may be the least of your concern.

Even for someone preferring free software, it may makes sense to have a non-free component on a Free Linux if there is no alternative, at least one doing that would have done 90% of the effort. For example in the past when ATI/AMD started to open their driver I bought AMD hardware while being OK if for some months or some years I would still have a proprietary driver until the free one gets better and replaces it.

But if someone is going to use Windows which is deeply closed source and on which both Nvidia and AMD drivers are closed source, that someone may do other choices.

Anyway Nvidia is known to create artificial market segmentation with drivers (which is why it looks unlikely they will release open ones on Linux as this would break this commercial strategy) so there may be other reasons to not prefer Nvidia on Windows but yes you're right that would not be about old hardware support on Windows.

[mdedetrich]

Anyway Nvidia is known to create artificial market segmentation with drivers (which is why it looks unlikely they will release open ones on Linux as this would break this commercial strategy) so there may be other reasons to not prefer Nvidia on Windows but yes you're right that would not be about old hardware support on Windows.

Actually NVidia doesn't want to create artificial segmentation for their drivers when it comes to OS (Linux/Windows), its that their hand is forced because of Linux being monolithic (i.e. it is expected all drivers are open source and sit in the tree). NVidia has no choice here and its not even in their interests to make NVidia Linux drivers have worse support (actually I would stipulate that the situation is the opposite, NVidia's biggest marketshare in Linux is HPC and any HPC computers that run old Linux versions cause headaches for NVidia, they don't want this problem).

It is true that NVidia does segment their drivers, but its done in other ways (i.e. for different GPU cards between consumer and pro). i.e. virtualization for GPU's is typically locked for higher end enterprise GPU cards (typically because the hardware virtualization market for GPU's is messed up and each card has to be individually tested, so this is NVidia's way of charging you extra money by forcing you to buy a card with higher markup).

[illwieckz]

Actually NVidia doesn't want to create artificial segmentation for their drivers when it comes to OS (Linux/Windows) […]

It is true that NVidia does segment their drivers, but its done in other ways (i.e. for different GPU cards between consumer and pro) […]

You may have missed something because when I talked about artificial market segmentation I talked about none of that, neither Linux/Windows segmentation, neither consumer/pro segmentation.

Nvidia is known to rely on their closeness to restrict the usage of their “general purpose graphic processing units” to not be general purpose ones.

For example recently they decided to implement some heuristics to detect mining tasks to prevent GPUs marketed for gaming to be used for mining. I don't care about mining, but I care about not having driver restricting this or that usage given the vendor agrees or not. That artificial market segmentation has also another very bad side effect: that destroys the second hand market for their own profit in a way that, when miners renew their hardware, their used hardware can't flow the second hand market for gamers or people doing workstation tasks. That also means products are sold to be used once and then thrown away, which is very bad on environment side of things. If Nvidia was opening their drivers and their firmwares, such artificial software white/blacklist would not be doable, this would completely defeat that kind of strategy. Because they promise to some customers (gamers) their GPUs can't be used for something else than those consumers want (gaming), they promise to those customers their drivers and firmwares will be closed, signed, etc.

You'll get a good summary there: https://www.youtube.com/watch?v=XfIibTBaoMM

[mdedetrich]

You may have missed something because when I talked about artificial market segmentation I talked about none of that, neither Linux/Windows segmentation, neither consumer/pro segmentation.

Nvidia is known to rely on their closeness to restrict the usage of their “general purpose graphic processing units” to not be general purpose ones.

For example recently they decided to implement some heuristics to detect mining tasks to prevent GPUs marketed for gaming to be used for mining. I don't care about mining, but I care about not having driver restricting this or that usage given the vendor agrees or not. That artificial market segmentation has also another very bad side effect: that destroys the second hand market for their own profit in a way that, when miners renew their hardware, their used hardware can't flow the second hand market for gamers or people doing workstation tasks. That also means products are sold to be used once and then thrown away, which is very bad on environment side of things. If Nvidia was opening their drivers and their firmwares, such artificial software white/blacklist would not be doable, this would completely defeat that kind of strategy. Because they promise to some customers (gamers) their GPUs can't be used for something else than those consumers want (gaming), they promise to those customers their drivers and firmwares will be closed, signed, etc.

You'll get a good summary there: https://www.youtube.com/watch?v=XfIibTBaoMM

I understand completely what you are saying. What I am saying is that none of whats related to this topic (ergo NVidia drivers having issues with older GPU's due to them to tied to outdated kernels) is due to artificial market segmentation so I have no idea why you are bringing this topic up.

[illwieckz]

I understand completely what you are saying. What I am saying is that none of whats related to this topic (ergo NVidia drivers having issues with older GPU's due to them to tied to outdated kernels) is due to artificial market segmentation so I have no idea why you are bringing this topic up.

This is because you answer to a comment about “you should not buy Nvidia” on a forum where the very vast majority of the public is expected to be both Linux users and people caring about software not being closed and software not arbitrarily limiting your usage. That's why the proposed arguments target that public. It looks like you wanted to defeat that statement with a use case that does not target such public (off topic argument being using proprietary drivers, whatever the brand, on an operating system that is not open source neither Linux), so I just recall the target of my comment was Linux users and people being concerned with open source and not being restricted by arbitrary limits in software. Then, if you rule out Windows, my first statement is still correct: “You should not buy Nvidia”, either either because their driver and then hardware support quickly being obsolete, either because of the way their development process don't fit with the open source ecosystem (meaning they get quickly out of interest), either there are other problems you will face (and I can list others I haven't listed at all)… This segmentation market example is only quoted to explain while Windows usage is out of topic here, which is the out of topic you brought.

Note that “NVidia drivers having issues with older GPU's due to them to tied to outdated kernels” is a very specific choice of naming the problem, others may say “NVidia drivers having issues with older GPU's due to NVidia not following the guidelines for kernel development”. It's a fact other vendors following the guidelines are not affected by such obsolescence, whatever the personal opinion we may have on such guidelines and their relevance.

Those different wordings are all about the problematic of Linux having to fit Nvidia development process or Nvidia having to fit Linux development process. And then, it's probably not wrong to assume the vast majority of the public there has interest on Linux before Nvidia (the need or the taste for Nvidia does not require Linux… and this website talks mostly about Linux and then Nvidia). You're the one who have expanded the topic out of Nvidia problems on an open source operating system for people who care about what they software limit or not. So I just recalled what was the primary public and its probable expectations. Even the initial article put the focus on the specific problem of Nvidia obsolescence on Linux. Outside of Linux the article only mentioned their general guidance but then, when talking about Linux, the article went in details about Linux-specific concerns like the Linux kernel itself, the Linux distributions, Xorg, Wayland or the alternative Linux native driver nouveau about its performance and feature limitations. The way Nvidia develop and distribute Windows driver and Windows users expectations or experience seem to be out of concern on the contrary.

[mdedetrich]

This is because you answer to a comment about “you should not buy Nvidia” on a forum where the very vast majority of the public is expected to be both Linux users and people caring about software not being closed and software not arbitrarily limiting your usage.

I was qualifying the statement saying that this issue is specific with Linux and not Windows and explaining why. In other words Linux is the elephant in the room here and its also constructive to have a look why

That's why the proposed arguments target that public. It looks like you wanted to defeat that statement with a use case that does not target such public (off topic argument being using proprietary drivers, whatever the brand, on an operating system that is not open source neither Linux), so I just recall the target of my comment was Linux users and people being concerned with open source and not being restricted by arbitrary limits in software.

Look tbh, pandering to people that sit in the Linux bubble and cannot handle criticism achieves nothing, and if people cannot take the heat in a public forum on Linux then they should reconsider why they are even participating in it.

Sticking your head in the sand and blaming everyone around you gets you nowhere fast.

So yeah, if you ignore Windows then the statement is correct but ignoring reality also resulted in the situation where you deliberately scapegoat everything onto NVidia rather than considering that "hmmm, maybe if the elephant in the room is the OS and not NVidia, then it might be that Linux also has a play to part"

Then, if you rule out Windows, my first statement is still correct: “You should not buy Nvidia”, either either because their driver and then hardware support quickly being obsolete, either because of the way their development process don't fit with the open source ecosystem (meaning they get quickly out of interest), either there are other problems you will face (and I can list others I haven't listed at all)… This segmentation market example is only quoted to explain while Windows usage is out of topic here, which is the out of topic you brought.

Actually its nothing intrinsically to do with open source (again you are misappropriating blame because of your world view). The core problem is actually of a technical one where Linux adamantly refuses any kind of ABI for drivers and has stuck to its guns in regards to being a monolithic Kernel. Drivers can still be open source and sit outside of the Kernel tree, or they can be closed, the core point here is about decoupling the graphics driver from the Linux kernel lifecycle. And to further reinforce the point that his has nothing to do with "open source", Linux already accepts programs that are closed source without any qualms as long as its in userspace, even drivers (i.e. FUSE), its just that unfortunately having a graphics driver in userspace hurts performance terribly.

And to be clear, this whole "Linux wants all drivers to sit in the tree" is also harming open source AMD graphics drivers as well, its just a problem that the Linux community conveniently ignore (or juts whines about and never gets solved). You have people crying on the forum why "Linux only has <1% market share on Steam", well its because no AAA game is going to market Linux when the capabilities of the graphics driver are tied to the Kernel (and its version) and due to how Linux distributions work, people can be running everything from the latest Kernel all the way down to Kernel 5-10 years old. So when a new gaming feature for graphics cards gets released (FSR, ray tracing, DLSS, VR etc etc), if the driver is open source then only the latest Kernels will support this feature (which is a fuked up situation for games specifically targeting those new features). Furthermore, since both games (and graphics drivers) are f**ken complicated, its typical for games and/or drivers to receive quick hotpatches 1-2 weeks after a games release. If the driver is decoupled from the kernel, this is really easy (i.e. you just download a new "package" which is your graphics driver) however with the Linux kernel you are probably waiting minimum for months on a new release (considering how fast distributions also pick up Kernel versions and/or backport it to older Linux Kernels if they even do so).

It can also ironically make support for older graphics drivers worse in some situations, because if the driver is decoupled from the Kernel and an improvement is done to older cards in a newer graphics driver release (which has happened), then everyone can benefit from the changes by just updating the graphics driver (even if they are on an older Kernel). With the way Linux works, you are relying on a backport which doesn't happen for multitude of reasons (a major one being that the code has changed so much that a backport is major work and also because its graphics drivers we are talking about, open source devs tend to care less about backporting these features to older Kernels because they assume not many peoeple care).

As you can see, these are all technical problems with how Linux is designed, its really got little to do with open source. Its also a good thing to point out that Windows/OSX also started as a monolithic kernel but moved to a hybrid kernel precisely to avoid these issues (and others).

Note that “NVidia drivers having issues with older GPU's due to them to tied to outdated kernels” is a very specific choice of naming the problem, others may say “NVidia drivers having issues with older GPU's due to NVidia not following the guidelines for kernel development”. It's a fact other vendors following the guidelines are not affected by such obsolescence, whatever the personal opinion we may have on such guidelines and their relevance.

Im sorry, you are deliberately misconstruing whats going on here. On a technical and factual level, the Kernel internally maintains no ABI and so its code constantly breaks. This means that if NVidia wanted to maintain every possible old Linux kernel out there, they would have to maintain a complicated matrix of graphics driver vs linux version (if we are talking about the blob) as well as retesting everything.

Mandating that NVidia does this is completely ridiculous and people that argue for this have no sense of normalcy/feasibility. If you are instead arguing that NVidia should completely open source their driver, as has been stated ad nauseam elsewhere, this isn't really possible for legal/IP reasons

Those different wordings are all about the problematic of Linux having to fit Nvidia development process or Nvidia having to fit Linux development process. And then, it's probably not wrong to assume the vast majority of the public there has interest on Linux before Nvidia (the need or the taste for Nvidia does not require Linux… and this website talks mostly about Linux and then Nvidia). You're the one who have expanded the topic out of Nvidia problems on an open source operating system for people who care about what they software limit or not. So I just recalled what was the primary public and its probable expectations. Even the initial article put the focus on the specific problem of Nvidia obsolescence on Linux. Outside of Linux the article only mentioned their general guidance but then, when talking about Linux, the article went in details about Linux-specific concerns like the Linux kernel itself, the Linux distributions, Xorg, Wayland or the alternative Linux native driver nouveau about its performance and feature limitations. The way Nvidia develop and distribute Windows driver and Windows users expectations or experience seem to be out of concern on the contrary.

Im sorry, when you say "different words" this to me is just an excuse that "I cannot accept facts/reality or a more nuanced/accurate analysis of whats going". And as I said before, if you cannot handle this then you (and Linux users like yourself) have far bigger problems (and believe it or not criticism does mean Linux gets better even if they don't like it).