DirectX 12 Ultimate and Vulkan Low-Overhead Draw Call Dispatch: Historical Paradigm Evolution and Next-Gen Architectural Projections

[Max Miller]

I’ve been benchmarking low-overhead dispatch across both DX12 and Vulkan, but the theoretical driver overhead numbers don’t always translate to what I’m seeing in-game at 1080p competitive refresh rates. When you’re pushing 240fps, that per-call CPU bottleneck becomes the actual limiter, not the GPU. How are you all structuring your draw call batches in practice—single-threaded command recording with a thread pool, or fully multi-queue submission? And more importantly, does the descriptor heap model actually pay off when you’re dealing with heavy shader permutations, or do you find the descriptor table reuse in older APIs still faster for certain titles?

I’m also curious about how the explicit multi-queue approach on RDNA hardware compares to what you’re getting on Intel’s hybrid scheduler. Are you seeing real gains from splitting compute, transfer, and graphics queues, or is the overhead of synchronization eating the benefit? And for anyone porting legacy deferred renderers to the new models—how hard was the migration away from immediate-mode state objects, and do you think descriptor heaps will eventually make descriptor tables fully obsolete, or is there a niche where tables still win?

"while (competitive) { aim_train(); review_demos(); improve(); }"

[Author]

Posts