Page 1 of 1

Neural Power vs. Old-School Calculation

Posted: Wed Feb 12, 2025 2:37 am
by mwyoung
Hardware: Threadripper 7970x, 128 Gbs Ram @ 6800 mhz

The AlphaZero vs. Stockfish 8 match was a pivotal event in the history of artificial intelligence and computer chess. It took place in 2017, when DeepMind’s AlphaZero a self-learning AI played against Stockfish 8. A top traditional chess engine. The results were astonishing and demonstrated the power of deep reinforcement learning. And AlphaZero won convincingly. It played 100 games against Stockfish 8 and won 28 times, drew 72 times, and lost 0 games.

In the AlphaZero vs. Stockfish 8 match, Stockfish 8 was run on a 64-thread, 1 GB RAM system, using a 64 core Intel Xeon processor.

This setup was far below Stockfish 8's potential, as modern chess engines typically require much more RAM (e.g., 32 GB or more) for optimal performance. Additionally, Stockfish 8 was not given an opening book or endgame tablebases, which further weakened its play compared to normal tournament conditions.

Meanwhile, AlphaZero used Google’s Tensor Processing Units (TPUs), which were specifically designed for deep learning, giving it an advantage in terms of computational power.

Some criticized the unfair conditions, arguing that Stockfish was not at its best (no opening book, no tablebases, limited hardware, and a fixed 1minute a move time control).

In this match we will give Stockfish 8 everything required to perform at its best playing today's top neural net chess engine.

At the same time we will be limiting the latest Stockfish neural net engine to 1 CPU and 1Gb of hash.

In the AlphaZero vs. Stockfish 8 match, Stockfish 8 searched approximately 70 million nodes per second. On this match hardware Stockfish 8 will be searching well over 100 million nodes per second. And at times reaching above 300 million nodes per second.

The new match conditions will remedy Stockfish 8's earlier match conditions when playing AlphaZero.

Match Conditions:
100 Games
Time Control: 60 moves in 60 minuets.
Stockfish 8 - 64 CPU, 64 Gb Ram,
Stockfish dev 18 - 1 CPU, 1 Gb Ram.
Opening Book of 10 moves played with reverse colors.
Top 10 seven man tablebases.

https://www.youtube.com/watch?v=WkiWIHjBxPA

Re: Neural Power vs. Old-School Calculation

Posted: Wed Feb 12, 2025 11:10 am
by Homayoun
Thank you very much for this article and other recent great constructive activities!
Best regards

Re: Neural Power vs. Old-School Calculation

Posted: Wed Feb 12, 2025 8:03 pm
by mwyoung
mwyoung wrote:
Wed Feb 12, 2025 2:37 am
Hardware: Threadripper 7970x, 128 Gbs Ram @ 6800 mhz

The AlphaZero vs. Stockfish 8 match was a pivotal event in the history of artificial intelligence and computer chess. It took place in 2017, when DeepMind’s AlphaZero a self-learning AI played against Stockfish 8. A top traditional chess engine. The results were astonishing and demonstrated the power of deep reinforcement learning. And AlphaZero won convincingly. It played 100 games against Stockfish 8 and won 28 times, drew 72 times, and lost 0 games.

In the AlphaZero vs. Stockfish 8 match, Stockfish 8 was run on a 64-thread, 1 GB RAM system, using a 64 core Intel Xeon processor.

This setup was far below Stockfish 8's potential, as modern chess engines typically require much more RAM (e.g., 32 GB or more) for optimal performance. Additionally, Stockfish 8 was not given an opening book or endgame tablebases, which further weakened its play compared to normal tournament conditions.

Meanwhile, AlphaZero used Google’s Tensor Processing Units (TPUs), which were specifically designed for deep learning, giving it an advantage in terms of computational power.

Some criticized the unfair conditions, arguing that Stockfish was not at its best (no opening book, no tablebases, limited hardware, and a fixed 1minute a move time control).

In this match we will give Stockfish 8 everything required to perform at its best playing today's top neural net chess engine.

At the same time we will be limiting the latest Stockfish neural net engine to 1 CPU and 1Gb of hash.

In the AlphaZero vs. Stockfish 8 match, Stockfish 8 searched approximately 70 million nodes per second. On this match hardware Stockfish 8 will be searching well over 100 million nodes per second. And at times reaching above 300 million nodes per second.

The new match conditions will remedy Stockfish 8's earlier match conditions when playing AlphaZero.

Match Conditions:
100 Games
Time Control: 60 moves in 60 minuets.
Stockfish 8 - 64 CPU, 64 Gb Ram,
Stockfish dev 18 - 1 CPU, 1 Gb Ram.
Opening Book of 10 moves played with reverse colors.
Top 10 seven man tablebases.

https://www.youtube.com/watch?v=WkiWIHjBxPA
https://www.youtube.com/watch?v=064MjZGHtwE

Re: Neural Power vs. Old-School Calculation

Posted: Mon Mar 17, 2025 8:08 pm
by mwyoung
Stockfish 18 has finish the match with Stockfish 8.

Playing with a 64 to 1 CPU handicap Stockfish 18 surpassed Alphazero's performance playing Stockfish 8. Even with the handicap, and playing a much stronger setup for Stockfish 8.

Code: Select all

Result:
--------------------------------------------------------------------------------------
  #  name                      games            wins           draws          losses   score    elo    +    -
  1. Stockfish dev 18 (1 CPU)    100       38 100.0%       58 100.0%        4 100.0%      67     49   27   26
  2. Stockfish 8 (64 CPU)        100        4 100.0%       58 100.0%       38 100.0%      33    -49   26   27

Cross table:
--------------------------------------------------------------------------------------
  #  name                         score   games         1         2
  1. Stockfish dev 18 (1 CPU)        67     100         x      67.0
  2. Stockfish 8 (64 CPU)            33     100      33.0         x

Tech:
--------------------------------------------------------------------------------------

Tech (average nodes, depths, time/m per move, others per game), counted for computing moves only, ignored moves with zero nodes:
  #  name                        nodes/m         NPS  depth/m   time/m    moves     time
  1. Stockfish dev 18 (1 CPU)    135421K     1899481     67.3     71.3     75.3   5365.6
  2. Stockfish 8 (64 CPU)      12078233K   160715876     53.0     75.2     75.0   5633.4
     all ---                    5952052K    83241753     60.2     73.2     75.1   5499.5
 

Re: Neural Power vs. Old-School Calculation

Posted: Mon Mar 24, 2025 3:04 pm
by maad59
Thank you very much for arranging this match, which is in my opinion essential and worthy.
Can you please upload the games?
We read a lot about the wonderful AlphaZero games, and I hope to see similar or better games by Stockfish 18.
Best Regards