On November 17, 2026, Natalie Vock posted version 2 of a 10-patch AMDGPU series proposing batched GPU virtual-memory updates and explicit synchronization. The changes target stalls in workloads with heavy sparse binding, including some games running through Proton, with the aim of reducing stutters. The Linux kernel mailing-list thread describes the proposal and its review discussion.

How the proposed AMDGPU update would work

Sparse binding maps and unmaps regions of GPU virtual memory as an application needs them. Natalie Vock says that synchronizing PRT mappings with other submissions can hold up graphics and compute work when a workload makes many sparse updates.

Vock’s series would add a virtual-memory update context to track mapping changes and pending updates. Its proposed AMDGPU_GEM_VA2 ioctl would let an application submit multiple mapping operations in one call and use multiple input and output synchronization objects, including timeline sync objects. That design is intended to coordinate mapping work with other GPU submissions more explicitly.

Vock expects the approach to reduce stuttering in applications that issue many sparse operations, particularly some games running under Proton. The proposal’s reach depends on the workload: frequent sparse-memory updates are the target.

The proposed benefit is tied to a specific workload: applications that make many sparse-memory updates.

This is separate from NeoTeo’s earlier coverage of a low-VRAM Radeon memory change, which addressed a different memory-management effort.

What the review discussion adds

Automated Sashiko review messages in the thread listed 15 potential issues, including concerns involving ioctl handling, fences, and rollback paths. Dave Airlie suggested the name VM_BIND for consistency with xe and nouveau.

On November 19, Igor’s Lab reported that the patches had not entered a released kernel by that date.