From this link, AlphaGo would take millennia to run in regular hardware.

They generated 29 million games for the final result, which means it's going to take me about 1700 years to replicate this.

Are these calculations correct?

  • 1
    $\begingroup$ It's not unreasonable to think they trained AlphaGo on (the equivalent of) thousands of parallel instances of commodity hardware-- the months of training time probably represent a few million dollars of cost to Google. Note that a trained AlphaGo can run reasonably on a single commodity machine. $\endgroup$
    – antlersoft
    Nov 30, 2017 at 19:34

1 Answer 1


Although the above statement holds important analogies to communicate the technical advances made by deep mind in the development of Alpha Go. It is inaccurate and should be taken skeptically.

Firstly, although Alpha go was trained on specialized hardware such as high end NVidia GPU's and custom google TPU's. It should be noted that it can run on a regular desktop although it won't be as powerful as the distributed version of Alpha Go. Additionally anyone with a laptop can access a similar amount of computing resources in the cloud with the touch of a button.

Although a version of Alpha Go indeed ran on 1202 CPUs and 176 high end GPU's as reported by Nature magazine. Which roughly translates to a system 1000 times as powerful as a commodity laptop. We need to consider other important factors such as moores law which postulates that we could have a computer as fast as 1202 CPU's in our pockets within 20 years and not several millennia.

Furthermore, the latest version of Alpha Go, Alpha Go Zero, trained on a single server with only 4 TPU's. Considering that the latest generation of smartphones such as iPhone X and Huawei's Kirin come packed with specialised AI chips. I will not be surprised if a similar reduction of form factor is achieved in commodity desktop computers once new models packed with AI chips are introduced.

I respect and acknowledge the technical achievements of Demis Hassabis and the deep mind team in developing a system as powerful as Alpha Go. However I believe analogies such as the one used above are inaccurate and mis-represent facts.

  • 2
    $\begingroup$ You confuse runtime and trainings hardware. Alpha Go Zero ran on 4 TPUs, but it was trained on 5000 TPUs. $\endgroup$ Jan 16, 2018 at 8:51
  • $\begingroup$ @Seth Simba so do you think alpha go zero would be faster and easier to train? $\endgroup$
    – paws
    Mar 17, 2018 at 10:41

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .