I reconstructed the full AlphaGo Zero network in Caffe: https://sjeng.org/dl/zero.prototxt
I did some performance measurements, with what should be state-of-the-art on consumer hardware: GTX 1080 Ti NVIDIA-Caffe + CUDA 9 + cuDNN 7 batch size = 8 Memory use is about ~2G. (It's much more for learning, the original minibatch size of 32 wouldn't fit on this card!) Running 2000 iterations takes 93 seconds. In the AlphaGo paper, they claim 0.4 seconds to do 1600 MCTS simulations, and they expand 1 node per visit (if I got it right) so that would be 1600 network evaluations as well, or 200 of my iterations. So it would take me ~9.3s to produce a self-play move, compared to 0.4s for them. I would like to extrapolate how long it will take to reproduce the research, but I think I'm missing how many GPUs are in each self-play worker (4 TPU or 64 GPU or ?), or perhaps the average length of the games. Let's say the latter is around 200 moves. They generated 29 million games for the final result, which means it's going to take me about 1700 years to replicate this. I initially estimated 7 years based on the reported 64 GPU vs 1 GPU, but this seems far worse. Did I miss anything in the calculations above, or was it really a *pile* of those 64 GPU machines? Because the performance on playing seems reasonable (you would be able to actually run the MCTS on a consumer machine, and hence end up with a strong program), I would be interested in setting up a distributed effort for this. But realistically there will be maybe 10 people joining, 80 if we're very lucky (looking at Stockfish numbers). That means it'd still take 20 to 170 years. Someone please tell me I missed a factor of 100 or more somewhere. I'd love to be wrong here. -- GCP _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
