Linux 7.3: Improved Performance When VRAM Is Exhausted
Original: Linux 7.3 improves performance when running out of vRAM
Why This Matters
Better VRAM overcommit handling in Linux 7.3 could significantly improve gaming on systems with limited GPU memory.
Kernel patches improving GPU VRAM overcommit performance have been merged upstream and are queued for Linux 7.3. The work, developed by a contributor at pixelcluster.dev, aims to make games more stable and playable even when VRAM usage exceeds physical GPU memory limits.
A developer at pixelcluster.dev has announced that kernel patches improving VRAM overcommit handling have been merged upstream and are queued for release in Linux 7.3. The patches build on earlier work focused on VRAM management for games.
When a game exceeds the physical VRAM available on a GPU, some memory must be evicted to CPU RAM. This creates a performance bottleneck because GPU access to CPU RAM is slower and must travel over the PCIe bus. On a PCIe 4.0 x16 connection, bandwidth is approximately 32 GiB/s — meaning a GPU can access a maximum of roughly 1,075 MiB per frame at 30 FPS. If more than 1 GiB of data must be fetched from evicted memory in a single frame, hitting 30 FPS becomes physically impossible.
However, the developer notes that not all memory accesses to CPU RAM are equally costly. GPU caches can amortize the high latency of PCI fetches through cache hits. Microbenchmarks on RDNA3 hardware show that when data fits within L2 cache, access latency is identical regardless of whether memory resides in VRAM or CPU RAM. The patches focus on making VRAM exhaustion a manageable performance issue rather than a cause of crashes or severe instability.