SIMT architecture is why GPUs are actually not that great for ray tracing. In a nontrivial scene, some rays will bounce more times than others, and that means that the whole parallel execution unit is tied up until every single ray terminates. (The other "threads" essentially execute NOP instructions in the meantime.)

Despite what Nvidia would have you believe, a lot of VFX rendering is done on CPUs, not GPUs, and this is one of the reasons why. (For example, last I heard, Pixar does all their render work on Intel Xeons.)