Support us on Patreon
PayPalMesa 26.0 has arrived as the latest new-feature release of open source graphics drivers, and there's a whole lot to be excited about in this one. For AMD GPU owners especially, as the RADV driver had a number of ray tracing performance improvements.
In the release announcement from developer Eric Engestrom they mentioned:
I'm happy to announce a new feature release, Mesa 26.0.0!
Mesa is the open source 3D graphics library, which includes implementations of Vulkan 1.4, OpenGL 4.6, OpenGL ES 3.2, OpenCL 3.0, and more APIs.
Since Mesa 25.3, RADV has seen significant raytracing performance improvements.
Its `RADV_DEBUG` environment variable has removed the following deprecated options: invariantgeom, nodynamicbounds, nongg_gs and splitfma. Use the driconf variables radv_invariant_geom, radv_no_dynamic_bounds, radv_disable_ngg_gs, and radv_split_fma instead.
RadeonSI switched to ACO by default for better and faster ISA code generation and better GPU performance and better compile times.
This release also introduces KosmicKrisp, a Vulkan to Metal layered driver for macOS.
Users can expect the usual flurry of improvements across all drivers and components, including these new extensions & features highlighted by their developers (in no particular order):
- VK_KHR_relaxed_block_layout on pvr
- VK_KHR_storage_buffer_storage_class on pvr
- VK_EXT_external_memory_acquire_unmodified on panvk
- VK_EXT_discard_rectangles on NVK
- VK_KHR_present_id, VK_KHR_present_id2 on HoneyKrisp
- VK_KHR_present_wait, VK_KHR_present_wait2 on HoneyKrisp
- VK_KHR_maintenance10 on ANV, NVK, RADV
- VK_EXT_shader_uniform_buffer_unsized_array on ANV, HK, NVK, RADV
- VK_EXT_device_memory_report on panvk
- VK_VALVE_video_encode_rgb_conversion on radv
- VK_EXT_custom_resolve on RADV
- GL_EXT_shader_pixel_local_storage on Panfrost v6+
- VK_EXT_image_drm_format_modifier on panvk/v7
- VK_KHR_sampler_ycbcr_conversion on panvk/v7
- sparseResidencyImage2D on panvk v10+
- sparseResidencyStandard2DBlockShape on panvk v10+
- VK_KHR_surface_maintenance1 promotion everywhere EXT is exposed (anv, hk, lvp, nvk, radv, tu, v3dv, vn)
- VK_KHR_swapchain_maintenance1 promotion everywhere EXT is exposed (anv, hk, lvp, nvk, radv, tu, v3dv, vn)
- VK_KHR_dynamic_rendering on PowerVR
- VK_EXT_multisampled_render_to_single_sampled on panvk
- VK_KHR_pipeline_binary on HoneyKrisp
- VK_KHR_incremental_present on pvr
- VK_KHR_xcb_surface on pvr
- VK_KHR_xlib_surface on pvr
- VK_KHR_robustness2 promotion everywhere EXT is exposed (panvk v10+, HoneyKrisp, hasvk, NVK, Turnip, lavapipe, venus)
If you want some background on the ray tracing improvements, developer Natalie Vock put up a blog post at the end of January with a load of technical info.
Going by their usual schedule, we should see the first bug-fix point release of Mesa 26.0.1 in around two weeks time, which you may want to wait for just in case any major problems snuck into this release. According to the current release calendar, the next new-feature release should be Mesa 26.1 due around May 6th. And the last bug-fix for the previous release Mesa 25.3.6 should be out on February 18th.
Article taken from GamingOnLinux.com.
See more from me
You can also find comments for this article on social media: 
Gerarderloper 18 hours ago
At least when looking at very heavy RT based games like many running UE5, or CP77.
I think the only reason AMD/Intel look ok in these games atm is because NVIDIA has that COLOSSAL 40-50% performance hit in MANY DX12 RT games. (probably still 3mnths until we see fixes be fully integrated)
Last edited by Gerarderloper on 11 Feb 2026 at 7:51 pm UTC
Quoting: GerarderloperAMD/Intel RT performance being quite far behind NVIDIA, and probably still behind.Yes, the $1000 more hardware is 1000$ worth better at raytracing, ofcourse.
At least when looking at very heavy RT based games like many running UE5, or CP77.
I think the only reason AMD/Intel look ok in these games atm is because NVIDIA has that COLOSSAL 40-50% performance hit in MANY DX12 RT games. (probably still 3mnths until we see fixes be fully integrated)
Quoting: SlayerTheChikkenIt's not quite that clear-cut, but it's not far off! It's awkward, because the top end AMD card, the RX9070, doesn't really have a direct equivalent Nvidia card - it's somewhere between a 5070ti and a 5080 (probably nearer the lower end), at least at 4K. I can pick up a RX9070 from scan.co.uk here in the UK for around £650 (which still, to me, feels like a crazy amount of money!). The 5070ti is around the £850. But the next jump up, the 5080 is about £1200, which is... unimaginable (to me).Quoting: GerarderloperAMD/Intel RT performance being quite far behind NVIDIA, and probably still behind.Yes, the $1000 more hardware is 1000$ worth better at raytracing, ofcourse.
At least when looking at very heavy RT based games like many running UE5, or CP77.
I think the only reason AMD/Intel look ok in these games atm is because NVIDIA has that COLOSSAL 40-50% performance hit in MANY DX12 RT games. (probably still 3mnths until we see fixes be fully integrated)
And then there's the £2750 you'll spend on a 5090. I mean, I wouldn't, but that's scan's price for anyone with money (and electricity) to burn!
But the point is moot for me anyway, on two levels:
a) I'll never buy an Nvidia card again, unless they fully open-source their drivers and get them merged into the kernel.
b) Ray-tracing is utterly underwhelming in every video I've seen using it.
TheLinuxPleb 2 hours ago
So far it seems that ray tracing implementations are just pretty shoddy and not worth the time. The ray traced reflections do work nicely in CP2077 where i think that the best practice is just to enable the reflections and leave everything else baked in settings.
Quoting: scaineb) Ray-tracing is utterly underwhelming in every video I've seen using it.Of course it is. In many ways it's similar to AI. The core technology is great and has valid uses, but the way it is presented for consumer use is nothing short of a scam. If you've ever done global illumination in something like blender, you're well aware of how high samples/pixel count needs to be to generate something that looks presentable. These numbers are so far away from anything that can be done in realtime, that you know something fishy is going on. Nvidia calls it "de-noising", but in reality it means that they have to "guess" 99.9% of the data. So basically upscaling on steroids. So the price of getting slightly better global illumination is a ton of artefacts. Of course that last part is never mentioned because you can't sell GPUs with that.
How to setup OpenMW for modern Morrowind on Linux / SteamOS and Steam Deck
How to install Hollow Knight: Silksong mods on Linux, SteamOS and Steam Deck