VFXVeteran
Did some math on modern GPU ray-tracing limits.
The GDDR raw latency (no cache) is about 300ns per write (AFAIR it's almost the same for all GDDRx).
Which gives us around 70GB/sec bandwidth for full random access.
At 1920x1080x30fps we get to 237Mb per frame, i.e. ~300 writes per pixel.
Which can be bottle-necked by ~20Grays/sec
Obviously if we divide between primary and secondary rays as 60/40 we can get to 50Grays (primary rays will be kinda "free") and if 80/20 it's 100Grays
So I will say with current GDDR memory 100Grays is probably the theoretical maximum.
Which is not bad for 1080p@30 = 1600samples per pixel but it's only 200 samples for 4k@60...