Linux VRAM management set for more improvements with kernel 7.3

great work, this is also coming to Bazzite very soon™️

Natalie Vock is doing an astonishing work. Awesome!

The opportunities provided by opensource stacks are great, as in, they allow all this to just happen, and to reach users far and wide. And this kind of improvements is likely to stack with others in many places; kernel, userspace, DE, graphic libraries, etc.

Open source is great for improvements.

Kernel 7.3 is packed with major performance improvements.

- Includes VRAM management (dmemcg-booster); game benchmark videos have shown a massive performance boost.
- Brings performance improvements for older processors and low-end ("potato") PCs.
- Addresses a resource-heavy security mitigation affecting Intel and AMD processors; the fix improves performance across the board—from pre-10th Gen chips to Intel Atoms (this is especially useful for PCs that didn't get Windows 11 support because they were older than 7th Gen).
- Offers SMP improvements to reduce latency and lag when closing games and browsers or clearing memory-intensive assets like maps and textures; it should also help prevent in-game stuttering and optimize core and thread management.
- Improves Vulkan support for Nvidia.
- Improves management of hybrid CPUs (efficiency and performance cores) in the latest Intel and AMD generations.
- Improves battery management on laptops.
- Includes hardware compatibility improvements for certain handheld PCs (minor tweaks).
- performance improvements in btrfs and ext4

First of all, congrats to her, that's hella impressive.

Now, I tried reading her blog post, it's way out of my league for me. Tried searching some key words, doesn't seem to be any mention of Intel / ANV or Nvidia / NVK, but AMD does get mentioned with " AMDGPU " hits though.

I'm unsure why any press article I've read so far (not to throw shade on anyone, I'm grateful these articles do exist) never mentioned that, unless I'm misunderstanding something, my understanding is that, at least so far, it would only be for AMD right ? Please correct me if I'm wrong.

But if I do understand it correctly, then it does seem like a pretty big omission (not on purpose mind you, I can't fault people for missing a mention that I myself am unable to fully comprehend her article) that can lead to confusion, people thinking that everyone will benefit from it, while so far I haven't seen a single mention of another GPU other than AMD.

Don't get me wrong here : I'm impressed by what she's doing and what she already has done. It's so much more detailed than what I'm able to understand (I got the gist of it basically). And it's awesome that a decent number of people will be able to benefit from it.

It's just that it may not be as extensive a reach (i.e not including Intel & NVidia) as people might think when reading about it thinking it would apply to them when in fact it would not. Though I guess, at least for ANV and NVK, given that they're open source, there could be ways it could, perhaps, in the future, be used for them. Still, pretty good news though for the AMD people !

Last edited by TriciaPearson on 20 Aug 2026 at 3:49 am UTC

Quoting: TriciaPearsonmy understanding is that, at least so far, it would only be for AMD right ?

I'm quite certain that the current production version in SteamOS is targeted at the Steam Deck and Steam Machine, both of which use amdgpu. But I believe that was the same with her previous work, which was initially targeted at the specific hardware, and had a broadened scope as part of the upstreaming effort. I expect the same will apply here.

On the Nvidia side, (practically) no one uses NVK in production yet, as the performance isn't there yet. So Nvidia users are unlikely to benefit unless the work can be adopted into their open source kernel modules, or Nvidia themselves adopt the features. But if Nova/NVK improve rapidly over the next year or two, things might be different. Look at what happened with RADV and AMDVLK.

Quoting: Phlebiac

Quoting: TriciaPearsonmy understanding is that, at least so far, it would only be for AMD right ?

On the Nvidia side, (practically) no one uses NVK in production yet, as the performance isn't there yet.

The reason I was mentioning NVK is because Valve's engineers, during a video on Adam [something] review of the Steam Machine, said they were collaborating with Nvidia to bring Nvidia GPU / driver support to SteamOS as soon as possible though very unlikely before the end of 2026, and they mentioned the words " open source " about 15 seconds before that, in that interview. There's a 0% chance Valve would use Nvidia's own proprietary drivers, which means they were talking about NVK. And there lies the tie with my mention of NVK and Valve / Natalie's work 😅

I know right now it's not up to par, but given Valve's track record of pushing for softwares and drivers and getting them way better than the proprietary drivers, I have no doubt they'll be part of a push to improve it / speed it up, considering they want it for their SteamOS as soon as it's able to have decent performance and good feature parity with proprietary drivers (since while Linux users favours AMD, generalized OS agnosticc gaming market still puts Nvidia's gaming GPU marketshare at 90% so if you want to convince people to swap you better have a good support for it).

(Edit : Also they were mentioning upstreaming it for other distros, where AMD isn't the only GPU brand to exist, regarding why I mentioned Nvidia / Intel :p)

Last edited by TriciaPearson on 20 Aug 2026 at 8:09 am UTC

Quoting: TriciaPearsonFirst of all, congrats to her, that's hella impressive.

Now, I tried reading her blog post, it's way out of my league for me. Tried searching some key words, doesn't seem to be any mention of Intel / ANV or Nvidia / NVK, but AMD does get mentioned with " AMDGPU " hits though.

I'm unsure why any press article I've read so far (not to throw shade on anyone, I'm grateful these articles do exist) never mentioned that, unless I'm misunderstanding something, my understanding is that, at least so far, it would only be for AMD right ? Please correct me if I'm wrong.

But if I do understand it correctly, then it does seem like a pretty big omission (not on purpose mind you, I can't fault people for missing a mention that I myself am unable to fully comprehend her article) that can lead to confusion, people thinking that everyone will benefit from it, while so far I haven't seen a single mention of another GPU other than AMD.

Don't get me wrong here : I'm impressed by what she's doing and what she already has done. It's so much more detailed than what I'm able to understand (I got the gist of it basically). And it's awesome that a decent number of people will be able to benefit from it.

It's just that it may not be as extensive a reach (i.e not including Intel & NVidia) as people might think when reading about it thinking it would apply to them when in fact it would not. Though I guess, at least for ANV and NVK, given that they're open source, there could be ways it could, perhaps, in the future, be used for them. Still, pretty good news though for the AMD people !

My (very limited) understanding is that most (all?) of what is needed for this patch is in the TTM kernel layer, that the article states is "the shared Linux GPU memory management layer". So I think this latest patches should work for all GPUs.

The also impressive previous work from the author to improve VRAM management, is also generic, but requires the GPU driver to support cgroups. amdgpu driver supports cgroups, so it works for AMD cards using this driver, the author mentions that it should also work on Intel GPUs but did not test it so no guarantee, and unfortunately the NVIDIA driver does not support cgroups, so you cannot get this big perf boost on NVIDIA GPUs (please reconsider buying NVIDIA GPUs, specially if you are targeting Linux).

Last edited by doragasu on 20 Aug 2026 at 8:09 am UTC

Quoting: doragasuMy (very limited) understanding is that most (all?) of what is needed for this patch is in the TTM kernel layer, that the article states is "the shared Linux GPU memory management layer". So I think this latest patches should work for all GPUs.

The also impressive previous work from the author to improve VRAM management, is also generic, but requires the GPU driver to support cgroups. amdgpu driver supports cgroups, so it works for AMD cards using this driver, the author mentions that it should also work on Intel GPUs but did not test it so no guarantee, and unfortunately the NVIDIA driver does not support cgroups, so you cannot get this big perf boost on NVIDIA GPUs (please reconsider buying NVIDIA GPUs, specially if you are targeting Linux).

Okay so I was partially wrong then and misunderstood some stuff, it's not that it's only work done for AMD GPUs (I was wrong about that), but that only they support it at the moment then (meaning I was still right about the fact lots of people still won't be able to enjoy it sadly), interesting description thanks !

As far as reconsidering buying Nvidia sadly you're one year and half too late, and I'm not changing anytime soon. I'm just hoping the DX12 performance issue with Descriptor Heap gets fixed before end of year hopefully on proprietary drivers, and then I'm hoping NVK will pick up some more *Steam* (he he he...) in 2027, considering that it already had a good increase in performance in games in 2026 so far. And like I said in my previous post, with Valve behind it needing to get it to a good place for SteamOS, late 2027 might just be NVK's shining moment.
And Nvidia's official Nova open source new rust kernel is also getting updated almost each major kernel release (or every 2-ish major kernel release) so there's that too.

Quoting: TriciaPearson

Quoting: doragasuMy (very limited) understanding is that most (all?) of what is needed for this patch is in the TTM kernel layer, that the article states is "the shared Linux GPU memory management layer". So I think this latest patches should work for all GPUs.

The also impressive previous work from the author to improve VRAM management, is also generic, but requires the GPU driver to support cgroups. amdgpu driver supports cgroups, so it works for AMD cards using this driver, the author mentions that it should also work on Intel GPUs but did not test it so no guarantee, and unfortunately the NVIDIA driver does not support cgroups, so you cannot get this big perf boost on NVIDIA GPUs (please reconsider buying NVIDIA GPUs, specially if you are targeting Linux).

Okay so I was partially wrong then and misunderstood some stuff, it's not that it's only work done for AMD GPUs (I was wrong about that), but that only they support it at the moment then (meaning I was still right about the fact lots of people still won't be able to enjoy it sadly), interesting description thanks !

As far as reconsidering buying Nvidia sadly you're one year and half too late, and I'm not changing anytime soon. I'm just hoping the DX12 performance issue with Descriptor Heap gets fixed before end of year hopefully on proprietary drivers, and then I'm hoping NVK will pick up some more *Steam* (he he he...) in 2027, considering that it already had a good increase in performance in games in 2026 so far. And like I said in my previous post, with Valve behind it needing to get it to a good place for SteamOS, late 2027 might just be NVK's shining moment.
And Nvidia's official Nova open source new rust kernel is also getting updated almost each major kernel release (or every 2-ish major kernel release) so there's that too.

Again I have very little knowledge about the topic, so I might be completely wrong. But my understanding is that for newer NVIDIA GPUs (starting from Turing), most of the driver runs in a microcontroller (the GSP) inside of the GPU itself, and the Linux driver is a shim layer wrapping the endpoints provided by the GSP firmware. And as the firmware is encrypted and signed, you cannot touch it and the things that the Linux driver can do are very limited. Then there's NVK that is part of Mesa and is a userspace implementation of Vulkan for NVIDIA GPUs. But this runs in userspace and I think it cannot provide some of the things required (like the aforementioned cgroup support). That requires kernel, and in the NVIDIA case most likely also firmware support, and I suspect only NVIDIA can do it (unless they also open source the firmware running in GSP and provide a way to run your own firmware, but I wouldn't count on that).

Quoting: doragasuNatalie Vock is doing an astonishing work. Awesome!

Very impressive indeed, and all of that while still being a student!

Nice to see my 4gb Rx 570 will still get performance improvements.