I have a Radeon RX 6900 XT (16 GB VRAM, originally released in 2021), and it's possible to run some lightweight models to have okayish performance and quality of output, but nothing I've tried has come anywhere close to the quality even of the models I can use for free from OpenCode Zen or the free tier of Openrouter. If you want to keep everything local on the same card I have, it requires putting up with a model that's noticeably worse in virtually every metric than what you can get for free elsewhere, and the GPUs this article are talking about are three times as old as mine.
With an extra 8gb of vram you could run qwen 3.8 27b pretty comfortably, which isn't quite as good as frontier models but definitely on par with free models on openrouter and whatnot.
Also you can use multiple GPUs at once, two of your GPUs could run qwen 27b very comfortably, and with great performance.
OpenRouter has Qwen 3.8 27B in its free tier right now. My point was that if a 5 year old GPU can't even compete with the free offerings, the 15 year old ones mentioned in the article aren't going to come close.
I think you could replicate this with a larger model on your device.
Overall with the right quantisations for both the model and KV cache you can get a lot of mileage out of this old hardware. Speed remains the main limitation, as IIRC I was getting ~26-30tps on a 7700S.
Opencode currently has Mimo 2.6 Flash Free with 5x that context. Like I said, it's not impossible to get something usable, it's just not in the same league as what you can get without paying anything or needing to spend effort "managing" to get it working.
I'm not saying there's no benefit to using local models. My point is still the same as before: you have to be willing to sacrifice both performance and quality even when just comparing to free models that are available today.
Yeah, but some prefer this if it means they own the infrastructure.
I spent some years in the pharma industry, where AI came really late, as there was (justified) concern that sensitive data might leak - even by accident. Eventually a solution was implemented - it was some kind of open model (they didn't say which) running on-premises.
You couldn't tell these people that this and that powerful model is free or inexpensive, because it's useless to them if it's running on someone else's computer i.e. the cloud.
Sure, that's what the second paragraph of my comment that you replied to is addressing. The tradeoff might be worthwhile for some, but it is a genuine tradeoff.
Either way, my initial reply was responding to the idea of these 15 year old GPUs as potential "capable processing units" for LLMS. My argument is that they are not particularly capable at the current moment in time.
Most of these older cards are lacking the physical hardware for fp8 or other lower precisions that most quantized models use. Or the memory to run models at higher precision.
I just bought a used Ayaneo 2 handheld, it has an old(er) mobile RDNA 2 GPU and I was blown away by how well this thing performed under Linux. Almost everything (that's not a recent AAA game) runs beautiful and a lot faster/smoother than it does under Windows. The experience has been so good that I'm considering switching my main pc (with a 9070XT) to Linux as well.
No doubt Timur contributed heavily to this given Valves Steamdeck (which uses a very similar but slower GPU).
Given the current hardware prices it's pretty awesome to see someone squeezing maximum performance out of old hardware!
I'm sure Valve probably still deserves the credit, since the Steam Deck uses a RDNA 2 gpu, but the article mentions that Timur specifically worked on gcn1.0 GPUs - e.g. radeon HD 7800/7900 from 2012 (!)
Those are pretty old but Im sure there are reasons. Valve has access to specs from their consumers through steam so maybe a large segment of the third world uses old GPUs for gaming.
My experience has been that for Linux you are always better buying older mid tier hardware because all of the issues and optimisations have already been worked out and those changes have flowed through to your distro so you aren’t waiting on kernel updates.
I've switched my gaming PC last year to Linux and it's been flawless.
Originally ran a 3070 which matches your "older mid tier" description and then upgraded to a 9070XT maybe 6 months after launch so not so old or mid-tier.
Not in my experience. The 16GB 9070 non-XT seems to have been released in March 2025. I purchased one in November 2025 [0]. I observed -and continue to observe- no significant difference in performance in graphics performance between Linux and Windows. I've also not noticed any significant difference in performance with regards to raytracing, but that might be because I have the non-XT variant, and/or it might be because I run Gentoo Linux, which probably has much newer versions of both Mesa and the kernel than most Linux distributions. I CBA to test either of these theories.
[0] Mad props to Gamers Nexus for publishing the video where they speculated that the time around Thanksgiving was going to be the "low" point for graphics card prices... so if you were thinking of buying an upgrade that was THE time to do so. They were totally correct.
So I can't comment on the raytraced experience because I don't bother with the raytracing, however I found AND worked out of the box on day one multiple times now. However RT benchmarks from January show it neck and neck.
Ubuntu LTS distros can need some time, yeah. A few months is not unheard of.
AMD typically starts adding basic kernel support before the hardware comes out, but full feature support is usually shipped shortly after hardware release. I suppose it's to avoid leaking specs before release?
But anyway if you want good support for new hardware, much better to run some sort of rolling release, like most Arch based distros.
Personally I run an Ubuntu LTS but with the latest mesa drivers via ppa, and am always at least a generation behind on the card. Much cheaper that way. Or at least it used to be...
I use it for a bit of everything: dev work (InteliJ), basic 3D modeling/printing (freeCAD, Blender, Orca), light video editing (OBS, Kdenlive), gaming (wine) ...
Could probably go with an Arch based distro, but the Ubuntu/Debian base is useful for the good compatibility.
Yes, raytracing works and it’s performant. I don’t know about 90% of Windows 11 performance but I am
too busy enjoying games to care about benchmarks.
Just to put things into perspective, the 9070XT is the third or fourth fastest gaming graphics card money can buy depending on the game.
The 5070Ti is $300 more than the 9070XT with similar performance, the 5080 could give you a good boost but costs even more, then you’ve got the 5090 which is a completely insane and unrealistically expensive purchase.
Nvidia typically has more of a performance penalty in Linux than AMD as well, so that narrows the gap further.
This idea that it must have wonderfully perfect raytracing performance has to be put in this perspective. The reality is that even if raytracing is half as fast in Linux as it is in Windows you’re still raytracing much faster than the median gamer on the market.
I can't launch Doom 2016, Brothers or Darksiders 2 on Windows. The screen blanks, if I unplug the HDMI cable and plug it in again I do sometimes get a signal. Ofcourse it works flawless under Linux.
Who would've thought that in 2026 gaming under Linux is actually the better option. Pretty awesome to see it work so well!
My sentiments exactly, and now with the popularity of the steam deck and other Linux handhelds, every sort of new utility related to gaming seems to have first-class Linux support.
About the only issue you could possibly have is anti-cheat and even that has a fair amount of Linux support.
> those changes have flowed through to your distro so you aren’t waiting on kernel updates
Unless you insist on running a server distro that's consistently obsolete by design (and if the notion of GPU comes into picture, that definitely shouldn't be the case), it shouldn't matter in practice. Every reputable desktop distro has a kernel/mesa stack that's up to date.
It still takes a few years for other people to find and patch the bugs.
~3 year old hardware seems to be optimal for linux in my experience. Ideally a thinkpad or something lots of linux devs have.
Buy newer and you'll be constantly finding bugs and having to apply workarounds, custom kernel builds, custom config lines, etc. just to make basic features work like the ability to adjust screen brightness.
Don't buy anything ARM or requiring a custom bootloader - thats a constant battle with custom builds needed of nearly everything.
There's still a real amount of time between someone creating a patch and it going in to a final release of the kernel and then coming out even to rolling release distros. It's not years like debian but it's much longer than when consumers first get their hands on the hardware.
I switched to Bazzite about 3 weeks ago and am amazed at how well everything "just works". Steam and Battlenet games all running. Nvidia 5070, AMD Rizen 7800x3d, recent gigabyte motherboard.
My gaming PC has a 9070 non-XT, runs Linux, and runs all but a small handful of games [0] just fine. I suspect you'll be happy with the switch.
I run games both new and old and one of my favorite guilty pleasures is looking at the Steam forum for a four, eight, or ten year old game that I've started playing because it has recently become popular again and reading the complaints from Windows users about how a driver update screwed up the game... whether because of glitchy or incorrect graphics or unavoidable crashes. Meanwhile, I'm cruising along on Proton with zero issues. :smug-face:
[0] ...that small handful includes those that go out of their way to be incompatible with Proton...
Even proprietary drivers can use Mesa. Most of Mesa is MIT-licensed.
It's just that Nvidia doesn't care much about Linux, and -as always- Nvidia ignores what everyone else is doing and does their own thing. Sometimes doing their own thing works out really well in the short run, but -long term- they always fall behind.
Yeah, Nvidia cares a lot if you are using compute (aka CUDA). If you are using their pure compute offerings, they are really top of the line on performance and tooling.
Though only with commercial distros. If you're using a fixed version of RedHat, SUSE or Ubuntu (which Canonical offers a commercial version), you're fine. Just hope that their support is long enough and NEVER EVER change kernel versions except patches!
If your GPU is older than 5 years, then you're in danger zone and you should have completely replaced your hardware 3 years ago and bought new Nvidia equipment now 10x more expensive!
The biggest issue was Nvidia's insistence on buffer management with EGLStreams instead of GBM, which broke a lot of desktops. But they forfeit that fight ~4 years ago.
EGLStreams actually was(is) a pretty good API, and I think most compositors would've had a much better time if they decided to implement that on top of GBM.
To summarize the difference between the two, with EGLStreams, you just request a buffer to draw into from a pool, then once you're done, you just submit it, thereby relinquishing control of it.
With GBM, you explicitly need to manage the lifetime and access of your buffers, making sure you don't leak or accidentally write it when the GPU/driver tries to, there's no clear ownership, and you have to handle surfaces for double/triple buffering.
It's a much lower level API, and much harder for drivers and app developers to get right, which was no doubt responsible for many years of buggy Wayland compositors - in fact, basically there wasn't a single correct implementation until Valve built Gamescope/their wlroots-based KWin fork and did a bunch of driver work.
They care about Linux, on their own distribution and licensing terms, NVidia even has their own distro.
Like most FOSS done by companies, what they don't care is GPL, which sadly will eventually be meaningless after boomer and Gen X devs that created the GNU movement no longer walk this plane.
There isn't a single embedded OS FOSS alternative to GNU/Linux that uses GPL, including Linux Foundation's own Zephyr.
Having two Mesa installations at once is how you summon dragons to fight. The FOSS Linux ecosystem doesn't really play well with multiple versions of libraries.
Except they already do (and always have). amdgpu is baked into mainline DRM, therefore you have better support than Nvidia, whose drivers are proprietary and taint the kernel on load.
Honestly, the folks writing inference drivers for Llama.cpp / GGML would sort of benefit from better compiler work like this.
In general, Valve's work has been exceptional and supplementing AMD's own ROCM/OpenCL and Vulkan team, they've gotten a lot of defaults right and people should work together with Valve to improve support for their chips.
Use as a dedicated GPU for encoding and decoding video.
Post processing like frame interpolation or superresolution. Use for GPGPU workloads. Run additional monitors independently. Use for GPU passthrough to virtual machines. Use as a backup GPU for troubleshooting. Use for test code without breaking the main GPU
For video decode if it supports the codec you want you're set, but video encode has definitely improved on more modern GPUs. Older GPUs work but you'll likely get worse quality for a given bitrate.
Assuming Valve funds this work, just buy their products. If possible, avoiding credit cards as the CC companies are trying to dictate policies about which games are acceptable and which aren't. I buy the occasional steam card with cash. I encourage everyone to shift a bit of their spending away from CC's in order to reduce the power of these companies to tell us how to live our own lives.
You can still buy steam gift cards at Best Buy if you're in North America, other similar retailers likely carry them in other countries. But Steam said they're shutting down the physical program at the end of the year because of scammers.
Yeah AI is extremely good at reverse engineering compared to humans. It's one of those things were there's a huge amount of tedious but not super complex work. I think if you can set up a good harness for the agent it can probably write complete drivers.
The most difficult part is likely that a lot of hardware is easily brickable if you do the wrong thing, so I'm not going to let it loose on e.g. my solar inverter.. even though their software is shit and I'd love to replace it. (Don't buy QCells.)
I understand the good will of the intent but... you have to understand that those driver code paths went thru massive QA over the years. On such complexe hardware, any modifications, even believed benign, can be disastrous for some/many software actually still 'in production'.
This is very risky: one guy QA over a massive amount of software with hardly a few remaining users able to report issues of with those GPUs...
Better have a slow working well tested driver, rather than a broken driver which is supposed to be faster...
That is a valid concern in theory, but in practice, improvements to drivers for old GPUs rarely break things. There are regression tests... and part of it might be Timor being very diligent in his work.
This is my point: you won't see them because the people able to report the breakage are rare to non-existent anymore, because most those who were able to do so did move to recent GPUs. The real "normal" users with old GPUs will be forced to move to newer hardware.
The "tests" are nowhere near the QA which has been done by the real stuff out there. Come on, this is border line hypocrit.
I am sorry, but I have to severely disagree: this has more a taste of planned obsolesence than actually improving old GPU performance.
I got a r9 285 in 2020 something while the crypto frenzy was still eating up all the gpus. It was 6 years old, got it for like 60 bucks if I remember correctly, best purchase ever, I got to play gta V at 60fps again.
Try to remember back to 2015 the graphics were not that bad. 4k performance was poor but for 1080p 60fps on pretty much anything.
4K is kind of pointless for gaming anyway, IMO. I've got a 32" 4K 165 Hz OLED screen and I much prefer 1080p at 165 Hz over 4K at <= 60 Hz and a pegged GPU. I don't really see pixels either way. 4K is nice for text, though!
4k works great when combined with FSR. On a 9070 GPU the frametime cost of FSR 4 is around 2 msec so you can just render the game at 1080P, upscale to 4k.
I usually play like this, it gives a very sharp image at 120-180W with a barely audible fan.
FSR 4/DLSS/Xess is something else than a simple spatial image based scaler though.
If you haven't tried machine learning based upscalers do give it a try! The newest generation (DLSS 4.5/FSR 4.1) are in general preferred by users over a native image since they provide a far superior way of native AA.
AI based frame-gen is also very good. 4k/120 Hz (1080p/60 FPS rendered) is very doable on mid-range cards these days.
Note: The GPU is an AMD 6800 XT, which is AFAIK more fiddly for AI upscaling. I'm not interested at all in frame generation, it does nothing for latency.
I now have a 2011 or 2012+ 27" iMac everywhere I need a computer, running Linux Mint Mate, and it works better and with less frustration that any new $2500+ Lenovo workstation laptop my employers give me.
93 comments
[ 0.12 ms ] story [ 44.9 ms ] threadHow long is this runway?
Also you can use multiple GPUs at once, two of your GPUs could run qwen 27b very comfortably, and with great performance.
With AMD the best you can do in the consumer market right now is an RX 7900 XTX which is about 960 GB/s.
The other day I managed to get a context of 195k for Qwen3.5-9b Q4_K_M using a llama.cpp fork that supports TurboQuant:
https://github.com/TheTom/llama-cpp-turboquant
I think you could replicate this with a larger model on your device.
Overall with the right quantisations for both the model and KV cache you can get a lot of mileage out of this old hardware. Speed remains the main limitation, as IIRC I was getting ~26-30tps on a 7700S.
I'm not saying there's no benefit to using local models. My point is still the same as before: you have to be willing to sacrifice both performance and quality even when just comparing to free models that are available today.
I spent some years in the pharma industry, where AI came really late, as there was (justified) concern that sensitive data might leak - even by accident. Eventually a solution was implemented - it was some kind of open model (they didn't say which) running on-premises.
You couldn't tell these people that this and that powerful model is free or inexpensive, because it's useless to them if it's running on someone else's computer i.e. the cloud.
Either way, my initial reply was responding to the idea of these 15 year old GPUs as potential "capable processing units" for LLMS. My argument is that they are not particularly capable at the current moment in time.
No doubt Timur contributed heavily to this given Valves Steamdeck (which uses a very similar but slower GPU).
Given the current hardware prices it's pretty awesome to see someone squeezing maximum performance out of old hardware!
Originally ran a 3070 which matches your "older mid tier" description and then upgraded to a 9070XT maybe 6 months after launch so not so old or mid-tier.
I can attest to this myself, my 9070 was a bit rough on launch but worked perfectly soon after.
Not in my experience. The 16GB 9070 non-XT seems to have been released in March 2025. I purchased one in November 2025 [0]. I observed -and continue to observe- no significant difference in performance in graphics performance between Linux and Windows. I've also not noticed any significant difference in performance with regards to raytracing, but that might be because I have the non-XT variant, and/or it might be because I run Gentoo Linux, which probably has much newer versions of both Mesa and the kernel than most Linux distributions. I CBA to test either of these theories.
[0] Mad props to Gamers Nexus for publishing the video where they speculated that the time around Thanksgiving was going to be the "low" point for graphics card prices... so if you were thinking of buying an upgrade that was THE time to do so. They were totally correct.
https://www.techspot.com/news/110999-new-benchmarks-show-lin...
With significant improvements still incoming.
https://www.gamingonlinux.com/2026/01/even-more-amd-ray-trac...
AMD typically starts adding basic kernel support before the hardware comes out, but full feature support is usually shipped shortly after hardware release. I suppose it's to avoid leaking specs before release?
But anyway if you want good support for new hardware, much better to run some sort of rolling release, like most Arch based distros.
Personally I run an Ubuntu LTS but with the latest mesa drivers via ppa, and am always at least a generation behind on the card. Much cheaper that way. Or at least it used to be...
Could probably go with an Arch based distro, but the Ubuntu/Debian base is useful for the good compatibility.
Just to put things into perspective, the 9070XT is the third or fourth fastest gaming graphics card money can buy depending on the game.
The 5070Ti is $300 more than the 9070XT with similar performance, the 5080 could give you a good boost but costs even more, then you’ve got the 5090 which is a completely insane and unrealistically expensive purchase.
Nvidia typically has more of a performance penalty in Linux than AMD as well, so that narrows the gap further.
This idea that it must have wonderfully perfect raytracing performance has to be put in this perspective. The reality is that even if raytracing is half as fast in Linux as it is in Windows you’re still raytracing much faster than the median gamer on the market.
I actually switched off Windows in a last-ditch attempt to get Indiana Jones and the Great Circle to stop crashing. To my surprise, it worked.
Never going back to Windows.
Who would've thought that in 2026 gaming under Linux is actually the better option. Pretty awesome to see it work so well!
About the only issue you could possibly have is anti-cheat and even that has a fair amount of Linux support.
Unless you insist on running a server distro that's consistently obsolete by design (and if the notion of GPU comes into picture, that definitely shouldn't be the case), it shouldn't matter in practice. Every reputable desktop distro has a kernel/mesa stack that's up to date.
~3 year old hardware seems to be optimal for linux in my experience. Ideally a thinkpad or something lots of linux devs have.
Buy newer and you'll be constantly finding bugs and having to apply workarounds, custom kernel builds, custom config lines, etc. just to make basic features work like the ability to adjust screen brightness.
Don't buy anything ARM or requiring a custom bootloader - thats a constant battle with custom builds needed of nearly everything.
I run games both new and old and one of my favorite guilty pleasures is looking at the Steam forum for a four, eight, or ten year old game that I've started playing because it has recently become popular again and reading the complaints from Windows users about how a driver update screwed up the game... whether because of glitchy or incorrect graphics or unavoidable crashes. Meanwhile, I'm cruising along on Proton with zero issues. :smug-face:
[0] ...that small handful includes those that go out of their way to be incompatible with Proton...
Nvidia doesn't, and their Vulkan stack underperforms on Linux quite significantly.
It's just that Nvidia doesn't care much about Linux, and -as always- Nvidia ignores what everyone else is doing and does their own thing. Sometimes doing their own thing works out really well in the short run, but -long term- they always fall behind.
Nvidia cares a lot about Linux. Just... on their terms.
If your GPU is older than 5 years, then you're in danger zone and you should have completely replaced your hardware 3 years ago and bought new Nvidia equipment now 10x more expensive!
The biggest issue was Nvidia's insistence on buffer management with EGLStreams instead of GBM, which broke a lot of desktops. But they forfeit that fight ~4 years ago.
To summarize the difference between the two, with EGLStreams, you just request a buffer to draw into from a pool, then once you're done, you just submit it, thereby relinquishing control of it.
With GBM, you explicitly need to manage the lifetime and access of your buffers, making sure you don't leak or accidentally write it when the GPU/driver tries to, there's no clear ownership, and you have to handle surfaces for double/triple buffering.
It's a much lower level API, and much harder for drivers and app developers to get right, which was no doubt responsible for many years of buggy Wayland compositors - in fact, basically there wasn't a single correct implementation until Valve built Gamescope/their wlroots-based KWin fork and did a bunch of driver work.
Like most FOSS done by companies, what they don't care is GPL, which sadly will eventually be meaningless after boomer and Gen X devs that created the GNU movement no longer walk this plane.
There isn't a single embedded OS FOSS alternative to GNU/Linux that uses GPL, including Linux Foundation's own Zephyr.
In general, Valve's work has been exceptional and supplementing AMD's own ROCM/OpenCL and Vulkan team, they've gotten a lot of defaults right and people should work together with Valve to improve support for their chips.
Use as a dedicated GPU for encoding and decoding video. Post processing like frame interpolation or superresolution. Use for GPGPU workloads. Run additional monitors independently. Use for GPU passthrough to virtual machines. Use as a backup GPU for troubleshooting. Use for test code without breaking the main GPU
Is Valve funding this work in particular? How can people donate?
+1 to buying Valve's products.
https://help.steampowered.com/en/faqs/view/78E3-7431-1E88-AD...
Perhaps we will even be able to reverse firmware blobs into open source alternatives?
The most difficult part is likely that a lot of hardware is easily brickable if you do the wrong thing, so I'm not going to let it loose on e.g. my solar inverter.. even though their software is shit and I'd love to replace it. (Don't buy QCells.)
This is very risky: one guy QA over a massive amount of software with hardly a few remaining users able to report issues of with those GPUs...
Better have a slow working well tested driver, rather than a broken driver which is supposed to be faster...
The "tests" are nowhere near the QA which has been done by the real stuff out there. Come on, this is border line hypocrit.
I am sorry, but I have to severely disagree: this has more a taste of planned obsolesence than actually improving old GPU performance.
Try to remember back to 2015 the graphics were not that bad. 4k performance was poor but for 1080p 60fps on pretty much anything.
I usually play like this, it gives a very sharp image at 120-180W with a barely audible fan.
If you haven't tried machine learning based upscalers do give it a try! The newest generation (DLSS 4.5/FSR 4.1) are in general preferred by users over a native image since they provide a far superior way of native AA.
AI based frame-gen is also very good. 4k/120 Hz (1080p/60 FPS rendered) is very doable on mid-range cards these days.