Existential Crisis Of A Linux GPU Developer

Key Takeaways

When I was first seriously getting the PC hardware, I want to say the current Nvidia generation was the I want to say it was the 500 series and then very shortly after the 600 series came out. That is the Kepler architecture which is what the 700 series and the Titan went on to use. I remember when the Titan came out and the pricing of that thing was such a big deal because that was a $1,000 $999 MSRP GPU. And look at where we are now. Look at how bad prices were, especially during co when there was all the, you know, crypto mining and GPU scalping and all that fun stuff where prices just shot through the roof. Things are cheaper than they were, but now you have inflation problems and a whole manner of other problems. So, prices aren't where they're supposed to be. Now, here's some fun. Do you want to feel old? Do you know when the Nvidia 600 series first released? You might have a number in your head. It's probably longer ago than you thought it was. 13 years ago in 2012 and the first Titan released a year later in 2013. Now, I think most people kind of understand that the 600 series is like an old series because it's like a it's not the like thousand numbering we have now. So, here's another fun one. Do you know how long ago the Nvidia 10 series came out? So that's the 1070, 1060, 1080, all of that fun stuff. Because there are still people that use these GPUs today, 2016, 9 years ago, and even today, there are developers like Faith Extran working on better support for these ancient GPUs. At the same time though, not really knowing why they're even bothering doing so because this is ancient hardware. People are not buying it new. The user base for these cards is only getting smaller. So like why even put in the effort? And that's not me saying this. This is them having like an existential crisis about why they're doing this work. Today I submitted Vulcan 1.2 performance for NVK on all Kepler GPUs. They're not conformant yet. There's a review period. As I said, Kepler is 600, 700, and the original Titan Nvidia GPUs. And NVK is Nvidia Vulcan. This is an open- source project for supporting Vulcan on Nvidia GPUs. It is a separate thing from Nvidia's proprietary drivers. Sometimes I wonder why I even bother putting time into old hardware. Sure, there's a lot of GTX 1060s out there. They're going to become e-waste one day. So, this isn't a perfect number, but we can see the Steam hardware survey. And right here, GTX 1060, currently 2.24% of the Steam user base. Obviously, this is not every single card out there in existence. But 2.24% of Steam is already a lot of people. That's already a lot of GPUs and puts it at number 12 of the most popular GPUs on Steam. Which, looking at these numbers here, this is a bit of a side tangent. Looking at these numbers, it makes me wonder why in the world we're seeing AAA games release in the state they're releasing, requiring crazy high-end cards when these are the 12 most popular GPUs. You're not seeing 4080s and 4090s on here. What are you even doing? But back to the post. Also, we can't reclock them with Novo and likely we'll never be able to do so. So they just become e-waste. If you plug one in and want to run Novo on it, it's just going to burn a lot of PCIe power plugs and render slower than the GPU you got for free with your old Intel CPU. If you have an Nvidia GPU on Linux, you're most likely using a proprietary drivers directly from Nvidia. And even though a GPU like this is quite old and is by no means fast by modern standards, it's still going to be in most cases faster than some old Intel CPU. maybe a modern one, not so much. But when it comes to the opensource space, which is what MDK is a part of, is what Novo is a part of, which is another part of that open-source Nvidia stack, which most distros do give you the option to use with this era of Nvidia GPUs. There is basically nothing that can be done because they moved how controlling the clock of the GPU is done and there's not an open-source method to deal with that. Reclocking is controlled by the firmware. There's no way to deal with the registers exposed directly to PCI and the firmware is signed and Nvidia hasn't given us firmwares that are capable of reclocking. On Turing Plus, we're using the GSP firmware, which is the same firmware blob they use in their Windows and Linux drivers, and it just handles everything for us. Now, Cheuring, those are your 10 series cards, and then everything after that also makes use of the GSP firmware. This is why there is even talk of open- source Nvidia drivers nowadays because without this, you can't really do anything. You're going to run it at the lowest clock possible and costs a lot of money to use. Oh, but some of them you can reclock. Yes, Kepler, which is a 13-year-old architecture at this point. Sure, maybe with enough effort we can make it sing, but it's so old and missing so many features that it can't run modern stuff anyway. So, yeah, maybe we can replicate a 2012 gaming experience in 2025. And the point is, there is certainly some fun to be had with like period accurate gaming. I know some people are a big fan of doing that with like 80s systems and 90s systems. And I'm sure at some point in the future, people are going to want to do the same with early 2000 systems. And this is neat. This is nice. And this is also niche. This is always going to be like a very sort of hobbyist kind of group. As for the people that actually have these GPUs, because they're the GPUs they have, this is an evershrinking group. And frankly, they're going to have a much better experience using Nvidia's official drivers than what you get from the open- source ones in most situations. A big part of it, though, is because I care about Linux users. There are a lot of people who still have one of those old cards and still want to use it. Maybe they can't afford a new GPU. And the one they got secondhand from the rich uncle 5 years ago is the best they have. There's a lot of those people running Linux. There's a limit as to how much my time I can really justify spending on ancient hardware, but I want to give them what I can and that is absolutely respectable. I have nothing but respect for that position. But at the same time, I understand this sort of I guess mental separation. Maybe that's not the right word. this like this this idea where you know it's important to help support this old hardware for the people who can't do better, but you also know that's a fairly small group and you kind of want to put your time into more things that help more people. Now, this idea of spending time on ancient hard reminds me of a recent thing I saw in the LKML. It's a bit different because that person was kind of just doing it because they wanted to, but this was related to the dropping of the Intel 486 support. This is hardware that nobody's realistically using with a modern system like a, you know, a 1060, a 660. You could still actually be using it. It's going to be slow, but it's not a Intel 486. And this person was working on drivers for that hardware in 2025. Basically, nobody cared about the hardware, but they felt like it was a sort of fun engineering challenge and wanted to just go and play around with it. And I think that's a good enough reason to want to go and write some software. Not everything you do has to be useful to every single person. Sometimes you might just want to do something for fun. And that's kind of this idea here. Another is that I really like the archaeology of it. Today I went on a deep dive down the rabbit hole of how Nvidia has handled reading from global constant memory over the years. Tesla and Fermy didn't have any magic here. Now the name Tesla's been reused a bunch of different times with Nvidia. I think they are literally selling hardware right now using the Tesla name. Considering its mention alongside Fermy, I assume we're talking about the micro architecture from 2006. This went from the Nvidia 8000 series, which is a numbering system they used to use, up to the Nvidia 300 series, which is when they were still using the GT naming. And I believe in the 300 series, that's when the GTX naming was kind of coined. Then Furmy, that was the Nvidia 400, the 500, and the low-end part of the 600. Then Kepler B added a magic texture instruction that fetched from arbitrary 64-bit addresses. Then on Maxwell, they unified a bunch of caches and added a non-coherent cache mode, which skips all the coherency checks and just grabs the first thing that matches. On Voltater, they reworked their coherency model, but kept the non-coherent mode only under a different name. I love learning about this stuff. Now Maxwell is the 700 mobile 800 and the 900 series and VA weird inclusion in this line here but that is the professional architecture post 10 series. The third big reason is maintenance. Even if that old hardware will never run well, it does need to work. And the old Nova stack is such a pain in the ass to work on and try to maintain. If we can make zinc plus NVK even remotely decent on old hardware, that'll be a huge win for users in terms of reliability and stability and a huge win for the Nova community in terms of support costs. In the end, my job is to be a steward of the Linux graphics stack to ensure that Linux graphics is good. Actually, sometimes that means deciding that my time is better spent on new features and hardware so that Linux can stay ahead of the curve and sometimes that means I spend two weeks reviewing community patches and fixing the last few Kepler bugs so we can breathe some new life into some old GPUs. I know for an absolute fact there are definitely people in my comments that are making use of this hardware or in some cases even older hardware. Maybe you have like a 400 series GPU and that's the thing that you have because it's the only thing that you have. I understand entirely where this person is coming from where it's like I know that this is important, but there's this other thing that's important as well. And you only have so much time in the day to actually work on the things you want to work on. You have to sleep at some point. You have to eat at some point. you probably have a job that you have to do as well and that's going to take up some time. So, you can't just work on every single thing. You have to decide what you want to put your time towards. But I do appreciate the fact that there are people out there who go and work on this older hardware. Even if not that many people are using it, even if it's just for the sake of like, you know, a handful of people that maybe don't even know they're doing the work, maybe don't even appreciate it. They just kind of, you know, use the software. But let me know what you use in the comment section down below. Do you have something modern? Do you have something old? Maybe have a modern system and then happen to have your old hardware lying around and, you know, have some fun with that. Maybe you buy like old ThinkPads and mess around with those. I don't know. I know there's a bunch of people that like to mess with random things. So, um, I'm very interested to see what you're personally using. So, if you like the video, go like the video. If you really like the video and you want to become one of these amazing people over here, check out the patron subcribe silly bar pay linked in the description down below. That's going to be it for me. And what was your first GPU? Actually, better yet, did you have a system before GPUs were a thing? [Music] something like this.