Yea, I think LLMs would be much less powerful if not for the inherent randomness in the GPU...
It is like a decompression program that results in a slightly different result each time it is run on the same compressed data. If you are lucky, the delta could end up with answers to all your questions..