Poker-AI.org

According to the alberta cfr paper, they used a 200 node computer with 24 cores per node. Would it be more efficient to use gpus?

Maybe not. Often the benefits of a GPU cannot be realised because there is insufficient memory on the GPU or the data transfer time to and from the GPU hold it up. I've hit this problem myself. I don't know if UoA had this problem.

the advantage of using gpu for calculating is that gpus do float calculations much faster, but for CFR algo in poker, it is mainly MCMC and integer calculation, so gpus cannot do much better than cpu.
if you are using tensorflow or pytorch or other deep learning libs, you can see that they put integet operations on CPU instead of GPU.

But GPU also allows to parallel computations more than CPU. In CFR algo we have a place, where as much you can parallel - is better. This also is not important, or i missed smth?