Lumen Global Illumination Software vs Hardware Ray Tracing Tradeoffs: Fundamental Mechanics and Deep Mathematical Modeling

Between competitive titles and my push toward 240+ FPS locked frame times, I keep hitting a wall deciding between Lumen’s software GI and hardware ray tracing for dynamic lighting. Here’s where I’m stuck: Lumen’s screen-space approach with the new surface cache cards is theoretically cheaper per frame, but the radiance cache updates seem to spike frame time variance during heavy scene transitions. Does anyone have hard data on how the temporal reprojection error compounds across frames when you’re chasing sub-4ms budgets? And for the hardware RT crowd—how are you handling the BVH build overhead on dynamic geometry without tanking your 1% lows?

I’m less interested in marketing specs than the actual math under the hood. Has anyone done a breakdown of the cost model comparing Lumen’s coarse-to-fine sampling against hardware RT’s traversal costs at 1440p? Specifically, where’s the crossover point in scene complexity where hardware RT actually becomes more frame-time-predictable? And does anyone understand the physics of how screen-space radiance handles grazing angles and occlusion quality versus true path tracing? Would love to hear real-world numbers from people running these in fast-paced competitive setups.