Valve Engineer Optimizes VRAM for 8GB and Lower GPUs
Natalie Vock, an independent contractor for Valve working on the company's Linux graphics driver team and primarily on the open-source RADV Vulkan driver for AMD GPUs, has developed a set of Linux kernel patches and user-space tools that address VRAM eviction problems on graphics cards with 8GB or less of dedicated video memory.
The work, detailed in a blog post Vock published on April 9, 2026, improves gaming performance by ensuring that foreground applications such as games receive priority access to fast VRAM while background tasks are forced to use system RAM first.
The issue Vock targeted is a longstanding behavior in the Linux kernel's graphics memory management, particularly with the AMDGPU driver and the TTM memory subsystem.
On systems with limited VRAM, multiple applications compete for GPU memory. When VRAM fills up, the kernel evicts data to GTT, a slower memory region backed by system RAM and accessed over the PCI bus.
Games often suffer the most because background processes, such as web browsers with multiple tabs, can claim VRAM space and push game assets into the slower region, resulting in stuttering and frame-time inconsistencies.
Vock's patches introduce support for the DRM device memory cgroup controller, or dmemcg, which allows the system to designate certain processes as protected and control eviction order.
Two user-space components accompany the kernel changes.
- The first, dmemcg-booster, is a systemd service that enables and manages the cgroup limits for foreground applications.
- The second, plasma-foreground-booster, integrates with KDE Plasma to automatically identify the active window, typically a fullscreen game, and apply the prioritization.
For users not running KDE Plasma, newer versions of Valve's Gamescope compositor can also leverage the kernel capabilities when combined with dmemcg-booster. The full set of six kernel patches modifies TTM allocation and eviction logic alongside the new cgroup support.
Vock co-developed the dmemcg controller with Maarten Lankhorst at Intel and Maxime Ripard at Red Hat. In her blog post she explained the core mechanism.
"Long story short, between these kernel and user-space patches it will ensure that the game running on your system has first dibs to the available dedicated video memory before any spilling/eviction comes to the system memory."
She also noted the practical outcome in testing. "That is, until now. Because I fixed it."
The patches target discrete AMD GPUs for now, with potential extension to Intel Xe graphics cards and an upstream patch already sent for the open-source Nouveau driver for NVIDIA hardware.
They do not apply to integrated graphics solutions such as the Steam Deck's APU, which uses shared system memory rather than dedicated VRAM. Early testing on an 8GB VRAM card showed clear improvements in titles such as Cyberpunk 2077 running through Steam Play on Linux, where game data remained in VRAM longer and reliance on slower GTT memory decreased substantially.
The changes are already available for testing on CachyOS, which ships the required kernel version 7.0rc7-2 or newer along with the dmemcg-booster and plasma-foreground-booster packages in its repositories.
Users of other Arch-based distributions can install the tools from the AUR and compile a patched kernel if needed. Broader adoption depends on upstream acceptance into the mainline Linux kernel and official KDE packaging.
Vock's contribution arrives as Valve continues to invest in Linux gaming infrastructure, including Proton and SteamOS, and amid renewed discussion around 8GB VRAM configurations in upcoming Steam Machines.
The patches provide a software-level workaround that lets existing low-VRAM hardware deliver more consistent performance without requiring new hardware.
