← All resources

Vulkan WORM

Auto-discovered

Talk Advanced

Timothy Lattis

Visit resource ↗

Description

Vulkan WORM is a technical talk by Timothy Lattis on the Neokineogfx channel presenting Write Once, Read Many as a frame-scoped GPU memory model for simpler Vulkan compute and graphics pipelines.

It covers AMD RDNA caches and Vulkan memory qualifiers such as coherent, readonly, and writeonly coherent; write-through that avoids stale non-coherent lines; one write phase then many reads after a VK event; block sizing across cache lines and compression tiles; aliasing without mid-frame reuse flushes; indirect dispatch and scanout edges; and rules favoring one command buffer and one submit per frame with events instead of heavy resource tracking.

It is for advanced Vulkan engine authors reducing barrier complexity. Prior sync and cache experience is expected.

Access is a free YouTube lecture. Related AMD GPU submission reading is mentioned verbally in the talk.

It is useful as a compact mental model for barrier-light, bind-once frames when tracking overhead dominates.