Kernel patches queued for Linux 7.3 trace the worst AMD GPU crashes under VRAM pressure to a lock ordering deadlock in the memory manager, not to the memory system itself, and the same fix makes running low on VRAM survivable instead of fatal.
Read the full story: Linux 7.3 Makes Running Out of GPU VRAM Survivable →
Transcript
Everyone knows what happens when a game asks for more VRAM than your card has. It crashes. Turns out that was never really a memory problem. Kernel work queued for Linux seven point three traced those crashes to a locking bug. When two things fight over the same memory at once, the kernel detects a deadlock and is supposed to retry. The graphics memory layer didn't retry. It just failed the submission. Then there's the display buffer, which has to sit in one unbroken block of VRAM. When memory is fragmented, making room for thirty two megabytes of pixels could evict four gigabytes of everything else. The fix throttles that, and lets applications tell the driver which buffers they actually care about. The result, on an eight gigabyte card running a whole gigabyte over budget: nineteen point six milliseconds a frame. Perfectly playable.