Shader Compilation Stutter and PSO (Pipeline State Object) Caching: Edge Cases, Failure Modes, and Root-Cause Debugging

I’ve been chasing down micro-stutters in competitive shooters and I’m fairly convinced shader compilation stalls are the culprit, but I want to separate the real bottleneck from the noise before I start throwing hardware at it. When I profile with PresentMon and RGP, I see frame spikes that line up with first-encounter shader stages, but the async compilation queue seems to drain fine on paper. My question is: how do you reliably distinguish between a true async compile bottleneck and a driver PSO cache miss causing a sync stall? Are you looking at the compilation queue depth, or is there a better telemetry signal?

Second, on the mitigation side—how much are you relying on precompiled shader caches versus runtime async compilation? I’ve tried baking in a pre-compile pass for every map, but I keep hitting edge cases where dynamic material variants spawn unexpected PSOs and blow past my cache budget. Do you build your pre-compilation set dynamically by instrumenting frames, or statically from a known asset list? And for the driver-level PSO caches, how many of you have found the per-game cache directories actually persist correctly across sessions, or do you see them getting nuked by driver updates? Curious what your real-world failure modes look like before I commit to an approach.