Name Version Release Date CUDA Toolkit Data Center Driver for Debian 12  615.71.09 September 9, 2026 13.4 Release Highlights: Release notes, supported GPUs and other documentation can be found at: https://docs.nvidia.com/datacenter/tesla/index.html Data Center Driver for Ubuntu 26.04  615.71.09 September 9, 2026 13.4 Release Highlights: Release notes, supported GPUs and other documentation can be found at: https://docs.nvidia.com/datacenter/tesla/index.html Linux-aarch64 (ARM64) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added libnvcuextend.so, an internal support library that enables NVIDIA SoC engines such as Deep Learning Accelerator (DLA) and Programmable Vision Accelerator (PVA) to be programmed using CUDA. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Linux x64 (AMD64/EM64T) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Linux x64 (AMD64/EM64T) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Data Center Driver for Ubuntu 26.04  615.71.09 September 9, 2026 13.4 Release Highlights: Release notes, supported GPUs and other documentation can be found at: https://docs.nvidia.com/datacenter/tesla/index.html Linux x64 (AMD64/EM64T) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Linux x64 (AMD64/EM64T) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Data Center Driver for Debian 12  615.71.09 September 9, 2026 13.4 Release Highlights: Release notes, supported GPUs and other documentation can be found at: https://docs.nvidia.com/datacenter/tesla/index.html Data Center Driver for Ubuntu 26.04  615.71.09 September 9, 2026 13.4 Release Highlights: Release notes, supported GPUs and other documentation can be found at: https://docs.nvidia.com/datacenter/tesla/index.html Linux-aarch64 (ARM64) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added libnvcuextend.so, an internal support library that enables NVIDIA SoC engines such as Deep Learning Accelerator (DLA) and Programmable Vision Accelerator (PVA) to be programmed using CUDA. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Linux x64 (AMD64/EM64T) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Linux-aarch64 (ARM64) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added libnvcuextend.so, an internal support library that enables NVIDIA SoC engines such as Deep Learning Accelerator (DLA) and Programmable Vision Accelerator (PVA) to be programmed using CUDA. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Data Center Driver for Ubuntu 26.04  615.71.09 September 9, 2026 13.4 Release Highlights: Release notes, supported GPUs and other documentation can be found at: https://docs.nvidia.com/datacenter/tesla/index.html Linux x64 (AMD64/EM64T) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Linux x64 (AMD64/EM64T) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Data Center Driver for Debian 12  615.71.09 September 9, 2026 13.4 Release Highlights: Release notes, supported GPUs and other documentation can be found at: https://docs.nvidia.com/datacenter/tesla/index.html Data Center Driver for Ubuntu 26.04  615.71.09 September 9, 2026 13.4 Release Highlights: Release notes, supported GPUs and other documentation can be found at: https://docs.nvidia.com/datacenter/tesla/index.html Linux-aarch64 (ARM64) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added libnvcuextend.so, an internal support library that enables NVIDIA SoC engines such as Deep Learning Accelerator (DLA) and Programmable Vision Accelerator (PVA) to be programmed using CUDA. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322 Linux x64 (AMD64/EM64T) Display Driver 615.71.09 September 9, 2026 Release Highlights: Added support for VK_NV_low_latency revision 2, to enable out-of-the-box Proton support for NVIDIA Reflex in Vulkan-native games which use NvLowLatencyVk.dll. Fixed a bug which prevented VK_NV_low_latency2 from reducing latency on swapchains created from VK_KHR_wayland_surface and VK_KHR_display surfaces. Fixed multiple startup and alt-tab hangs that occurred with NVIDIA Smooth Motion enabled. Fixed a bug that caused a file descriptor leak when the Vulkan loader unloads libGLX_nvidia.so.0 after a call to vkDestroyInstance. Fixed an Xid 109 error when loading PRAGMATA on Wayland. Added support for cgroups-based memory partitioning. Added a new token, RmDisableDisplayGlitchPerfLimit, to the NVreg_RegistryDwords nvidia.ko kernel module parameter. This may reduce idle power consumption in some multi-monitor configurations, at the risk of changing memory clocks while display is using the memory, potentially resulting in momentary display glitches. Example usage: modprobe nvidia "NVreg_RegistryDwords=RmDisableDisplayGlitchPerfLimit=1" Fixed a bug that could cause corruption when display output scaling is used on Blackwell GPUs. Fixed a bug that caused resume from suspend to fail if the system was previously hibernated: https://github.com/NVIDIA/open-gpu-kernel-modules/pull/1199 Fixed a rare correctness issue on Blackwell GPUs affecting kernels with two or more nested levels of thread divergence, where reconvergence could fail. Kernels with only a single level of divergence are unaffected, as are applications already built with NVCC 13.2.2 or newer. This fix guarantees correct execution, though it may come with a modest performance cost. Recompiling with NVCC 13.2.2 or newer will resolve the possible slowdown. If you are unable to recompile your application for any reason, the runtime behavior of this fix is controllable on a per-process-launch basis through the environment variable CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE as follows: CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DEFAULT Default behavior: attempt JIT recompilation for 15 seconds, runtime patch remaining code CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=FORCED Re-finalize every eligible module, ignoring the time budget CUDA_SELECTIVE_DEVICE_CODE_RECOMPILE=DISABLED Driver patching only, never re-finalize. Enable support for the G-SYNC Pulsar feature in supported monitors https://www.nvidia.com/en-us/geforce/news/g-sync-pulsar-gaming-monitors-available-january-7-2026/ https://forums.developer.nvidia.com/t/gsync-pulsar-planned-for-linux/358322