NVIDIA 615.71.09 brings Reflex to Vulkan games under Proton and cgroups memory partitioning
On September 9, 2026, NVIDIA released the 615.71.09 Linux driver, enabling Reflex for native Vulkan games via Proton and adding cgroups-based memory partitioning to the open kernel module. Linux gamers and teams virtualizing GPUs gain two capabilities they had been missing compared to Windows.
September 9, 2026. NVIDIA released the 615.71.09 Linux driver, and two additions in particular stand out. The first is VK_NV_low_latency revision 2 support, which enables Reflex for native Vulkan games under Proton. The second is cgroups-based memory partitioning in the open kernel module, long requested by anyone running GPU workloads in containers. Why it matters: these two additions close gaps that still separated the Linux desktop from its Windows counterpart.
Reflex under Proton: latency becomes a Linux topic
Reflex is NVIDIA’s system-latency reduction technology, historically a Windows-stack exclusive. On Windows, games that want minimal latency rely on a low-latency Vulkan library shipped by NVIDIA. The 615.71.09 Linux driver now supports VK_NV_low_latency revision 2, which lets Proton use that same low-latency Vulkan library without specific adaptation — “out-of-the-box”, in Phoronix’s words.
In practice, a native Vulkan title run through Proton on Steam can now benefit from Reflex’s low-latency mode without the developer writing a single line of Linux-specific code. It is a change of mindset: latency reduction, long treated as a refinement reserved for competitive Windows gaming, becomes a standard building block of the Linux graphics stack.
Latency here is not a vanity metric. Reflex shortens the pipeline from mouse input to the frame actually displayed on screen — what competitive players call input-to-photon latency. On Windows, that low-latency path has been a selling point for years. Bringing it to Proton matters because Steam Play has made Linux a credible gaming platform, and until now the latency story was the one place a Linux box still gave ground to a Windows one.
The other notable Vulkan addition is VK_EXT_cluster_acceleration_structure, an extension aimed at engines that group acceleration structures for ray tracing. Less visible to end users, it signals that NVIDIA keeps widening its Vulkan surface on Linux beyond the bare minimum required for gaming.
The open kernel module gains memory partitioning
The second important piece concerns NVIDIA’s open kernel module — the driver variant whose code is published and progressively integrated into distributions. It now supports cgroups-based memory partitioning, a capability operators have been waiting for to cleanly isolate GPU memory across multiple containers or processes.
Without this building block, GPU memory is largely shared among the workloads on a machine, and it is hard to guarantee that one container will not spill over into another’s reservations. With cgroups partitioning, isolation follows the model already in place for CPU and system memory — a container is assigned a slice of GPU memory, and the kernel enforces it.
It is a meaningful convergence point with the Kubernetes ecosystem, where GPU scheduling and accounting increasingly go through standard resource-control mechanisms. The open kernel module catches up here on functionality expected of a modern Linux driver.
For operators running mixed inference and training on the same node — or sharing one large GPU across several tenants — this closes a concrete gap. Until now, the open kernel module offered no clean way to bound GPU memory per cgroup, so a single runaway workload could evict its neighbours. That isolation is now enforced by the same kernel primitives the rest of the stack already trusts.
A Linux driver playing on three fronts
Context helps measure this release’s reach. NVIDIA’s Linux graphics stack moves on three distinct fronts. The proprietary driver remains the reference for performance and features. The open kernel module, published in 2022 and now the default on several distributions for recent GPUs, exposes the same kernel code under an open license. Finally, the fully open-source Nova/NVK effort — the kernel driver and the Vulkan driver inside Mesa — aims, over time, at a community-maintained driver with no proprietary blob.
615.71.09 scores on the second front. cgroups-based memory partitioning was one of the last functions the open module still had to borrow from the proprietary driver to be credible in the data center. Its arrival pulls the open module closer to the standard GPU orchestration mechanisms that Kubernetes and container runtimes increasingly expect, as inference and training workloads densify onto shared machines.
Alongside it, the VK_EXT_cluster_acceleration_structure extension shows NVIDIA widening its Vulkan surface at the driver level rather than retreating into private extensions. For engine developers, that is a stability signal on a rendering path that matures with every release.
The fixes that affect day-to-day use
Beyond the foundational additions, 615.71.09 fixes several defects that annoyed real users:
- End of start-up and alt-tab hangs with Smooth Motion — NVIDIA Smooth Motion users hit freezes in both situations, now fixed;
- Fix for an Xid error when loading PRAGMATA under Wayland — a Wayland-specific issue that grew more relevant as distributions switch to the protocol by default;
- New
RmDisableDisplayGlitchPerfLimitkernel option — meant to reduce idle power draw in some multi-monitor configurations, a sore point for workstations and graphics stations.
Taken together, they describe a driver that no longer merely tracks Windows, but treats Linux as a first-class platform — with its own priorities around latency, power, and resource isolation. That shift has been years in the making, and it is now visible in release notes rather than marketing slide decks.
What users should do
The right move depends on the machine’s profile.
- Gaming rig under Proton — update to 615.71.09 if you play native Vulkan titles where latency matters, and confirm Reflex engages through the low-latency Vulkan library;
- Multi-monitor workstation — the idle-power fix and the Wayland corrections justify the update, especially on distributions that default to Wayland;
- GPU server running containers — the open kernel module’s cgroups memory partitioning directly affects you if you virtualize or partition GPUs; test the release on one node before rolling it out;
- Stable, conservative fleet — if none of the fixed defects hit you, nothing forces an immediate move, but do not let this release sit for more than one validation cycle.
Downloads and detailed release notes are available on the NVIDIA portal.
Verdict
NVIDIA 615.71.09 is more than a routine driver. It closes two long-standing gaps — Reflex for native Vulkan games under Proton, and cgroups-based memory partitioning in the open kernel module — that concretely separated Linux from Windows for two distinct audiences: latency-sensitive gamers, and operators who partition GPUs in containers.
If you game on Linux via Proton, this release is worth an immediate update for native Vulkan titles: it is the first time Reflex engages without specific adaptation, and latency is a measurable gain, not a cosmetic one.
If you run GPU workloads in containers, the open kernel module’s cgroups memory partitioning is the missing piece for cleanly isolating GPU memory. Test it on a validation node, then roll it out to clusters where GPU workload co-location already costs you resource incidents.