Great topic. The core issue historically was that runtime shader compilation forced the CPU to compile GLSL/HLSL on the fly, and whenever a new material, lighting combo, or post-process hit the pipeline, you’d get a hard stall while the driver fed the compiler. The shift to DX11’s PSO model was the real turning point—bundling all pipeline state (rasterizer state, blend state, depth-stencil, shader bindings) into a single object meant the driver could validate once and then hand the GPU a ready-to-go state, dramatically cutting per-draw-call CPU overhead and smoothing out frame times. DX12/Vulkan took it further with explicit resource binding and PSO caching to disk, so compiled states persist across sessions instead of recompiling every launch. If you’ve noticed stutter in modern titles, it’s often because the engine still falls back to on-demand compilation for dynamic materials (think destructibles or shader variants from LOD changes), and the cache miss on first encounter is exactly what you feel as a hitch.
Looking ahead, the architecture is moving toward async and mesh shader-driven compilation. AMD’s mesh shader pipelines and NVIDIA’s RT Core integration mean the compiler has to handle far more variable state, which is why engines like Unreal 5 lean into shader prewarming and async compute to off
"Tekken 8 Tekken King | Leverless Arcade Controller (Qanba Obsidian Mod)"
