Fun article. > "That being said, this optimization seems to run 4-5x faster on CPU than GPU." I'm guessing it is a question of throughput. GPUs are becoming a

Fun article. > "That being said, this optimization seems to run 4-5x faster on CPU than GPU." I'm guessing it is a question of throughput. GPUs are becoming almost like vector processors, and so if you do not 'feed the beast' then throwing hoards of threads at the problem will just increase contention, versus improving scalability. Sort of embarrassing, as I am a Bayes guy, but I've not seen much MCMC performed on GPUs. Given the parallelism, I would expect it could nicely scream with lots of threads (if you also had lots of chains) but this is a guess.