Shader Compilation Stutter and PSO (Pipeline State Object) Caching: Advanced Implementation and Real-World Architecture

[Mason Collins]

Been chasing shader compilation stutter through basically every title I play competitively, and I’m genuinely torn between approaches. Some devs are leaning on explicit pre-compilation baked into their asset pipelines, others are going full async with D3D12_SHADER_COMPILER_THREADING or Vulkan’s async compute-based compilation, and then there’s the whole driver-managed PSO cache question. My biggest frustration is that driver caches on Windows seem to fight each other across GPU vendors—NVIDIA’s cache vs AMD’s, plus the OS-level shader cache—so a build that’s smooth on one rig jitters on another. For anyone actually shipping async compilation in production, what’s your real-world strategy for the compilation queue itself? Are you using a priority-based system where you front-load the shaders for the current map and throttle background compiles, or is there a smarter scheduling model? And do you trust the driver cache or roll your own persistent cache keyed off shader source hashes?

On the engine side, I’ve seen people implement their own PSO cache with versioned keys tied to driver versions and hardware, but the invalidation logic always feels hacky. How are you handling cache poisoning from driver updates without nuking performance gains every month? And for benchmarks, are you measuring first-frame latency in specific hotspots, or is there a more reliable methodology than just watching FPS graphs? Would love to hear from anyone who’s actually profiled this with PIX or ROCt and has hard numbers.

[Author]

Posts