Mesa 26.2 NVK Driver Adds DLSS Support, Hidden Behind Experimental Flag

Q: Is NVIDIA's DLSS support a gift to Linux gaming or a Trojan horse for its proprietary ecosystem?

It is both, but right now it looks more like a practical gift with clear proprietary strings attached. For Linux gaming, the immediate benefit is real. Mesa’s NVK driver can now run DLSS through the new `VK_NVX_binary_import` extension, narrowing one of the biggest gaps between NVIDIA’s open-source Vulkan stack and its proprietary driver. That matters because DLSS remains one of the most valuable gaming features on RTX hardware, especially for high-resolution and ray-traced games. Its arrival in Mesa 26.2 also strengthens the case for using the open-source stack with Proton, DXVK and VKD3D-Proton. The catch is that this is not an open DLSS implementation. NVK is importing and executing NVIDIA’s prebuilt CUDA binaries, specifically CuBIN files containing architecture-specific SASS. That means the feature depends on NVIDIA’s closed code and on binaries matching the user’s GPU. NVK still lacks the PTX-to-bytecode path available in NVIDIA’s proprietary driver, so it cannot offer the same flexibility or independence. That is why the “Trojan horse” argument has force. The open-source driver gains a flagship feature, but only by becoming a host for NVIDIA’s proprietary software moat. Control over the core technology, compatibility and future support remains with NVIDIA, not the Mesa community. So this is a win for users, not a liberation from lock-in. Linux gamers get more choice in drivers, but not more control over DLSS itself.

Q: Could this breakthrough run NVIDIA’s AI ecosystem on open-source drivers, not just DLSS?

Not yet. This NVK milestone is important, but it is a narrow bridge to specific NVIDIA AI features packaged as precompiled GPU binaries, not a path to NVIDIA’s full AI stack on open-source drivers. What changed is that NVK can now use `VK_NVX_binary_import` to load NVIDIA CuBIN files, which contain GPU-native SASS code. That is enough for DLSS because NVIDIA ships the needed prebuilt CUDA kernels for supported GPUs. In practical terms, Mesa’s open-source Vulkan driver can now execute those binaries, so DLSS becomes possible without the proprietary Vulkan driver. The limit is the missing PTX-to-machine-code path. NVIDIA’s broader AI ecosystem depends heavily on CUDA’s full runtime, compiler chain, and libraries such as cuDNN, TensorRT, NCCL, and cuBLAS. Those workloads often rely on PTX JIT compilation, forward compatibility, and deep integration with proprietary user-space components. NVK does not currently provide that. It can run compatible precompiled binaries, but it cannot generally replace the CUDA software stack. So this breakthrough could help with other NVIDIA features that are similarly delivered as fixed, architecture-matched binaries and exposed through graphics or compute interfaces NVK can import. But it does not mean open-source Mesa drivers can suddenly run arbitrary CUDA applications or the wider NVIDIA AI ecosystem. It is best seen as a targeted compatibility win for DLSS and possibly similar packaged features, not a general open-source CUDA breakthrough.

3 articles · Updated · Phoronix · Jun 19

Code merged into Mesa 26.2 now lets the open-source NVK Vulkan driver run DLSS-enabled games on NVIDIA GPUs.
VK_NVX_binary_import makes that possible by importing NVIDIA CuBIN CUDA binaries, which DLSS relies on for execution through NVK.
Known bugs keep the feature gated behind the NVK_EXPERIMENTAL=dlss environment variable rather than enabled by default.
Support also depends on compatible GPU bytecode being available, because NVK still lacks the proprietary driver’s PTX-to-bytecode path and cannot yet translate PTX into Mesa’s NIR.
Mesa 26.2 is due to ship as a stable release in August, making this an early but notable step for NVIDIA gaming support in the open-source stack.