Julia uses a task local rng by default so we shouldn’t be running into the same issue cherno did.
The 22s image was 1264x761 with 32 samples per pixel and a max depth of 8. My benchmark is 1920x 1080 with 1000 samples per pixel and a 16 max depth.
At best cherno’s improved program would have run in 22 * 1920/1264 * 1080/761 * 1000/32 = 1482s.
Cherno’s pc seems a lot better than mine as well.