[IMPORTANT] AlphaZero vs. Stockfish Simulations Settings Are All Possibly Flawed!

[Cscuile]

TLDR: All calculations are based upon the information given by Deepmind on the NPS of Stockfish 64 CPU and AlphaZero 4 TPU. 70,000,000 NPS and 80,000 NPS. There is evidence to assume that these numbers are an average through the entirety of the game and NOT solely from the starting position of chess.

This is significant as it affects the draw rate of the AlphaZero-Stockfish 8 simulations. As of right now Leela and SF8 are drawing quite a bit more than A0 did against SF8. With the updated settings, we may see Lc0 win more against SF8. 



This will provide evidence based on if Leela has or has not surpassed/matched AlphaZero. 

In order to re-calibrate the settings we will need to do either one of two things. 

• Create a binary for Lc0 that runs on TPUs. 
  
• Estimate the NPS of A0 4TPU (At the starting position) based on what 64 CPU achieves.

Current Settings: 
4,200,000,000 NPM for Stockfish 8
4,800,000 NPM for Leela

Rough Estimates of the NPS at the starting position of chess. 

• Nodes per Second of 64CPU (Starting) = 44,000,000 NPS
• 44,000,000 / 70,000,000 (NPS DeepMind) = 0.6286 Difference Ratio
  
• 0.6286 * 80,000 = 50,286 NPS
  
• 44,000,000 * 60 Seconds = 2,640,000,000 Nodes per Move for Stockfish 8
• 50,286 * 60 Seconds = 3,017,160 Nodes per Move for Leela