Current TCEC Cup

[MikeGL]

LCO has a long way to go.

Lco would probably be on par with Komodo and Houdini( She won 1v1 on Komodo in cccc).She is already somewhere between top 2- 4. Her score 4.5-2.5 means , -100 elo, you will get similar result between SF vs K or H.

Why Stockfish is so strong?. Let us estimate the inveatment in SF project from fishtest data alone from 5 years ago. More than 1 billion games have been tested, 30,000+ patches, and

1400 years of CPU years of computing power have been used.( It means if you run 1 ghz cpu for 1 core computer, you will need 1400 years to complete these tests).

For estimate of electricity usage alone( not to mention hardware cost and human resources), I estimated that $ 4,000,000( 4 million have been spent in Stockfish project).( I estimate 1 core computer , 60 W, cost- 10 cents per day).

Yes, a long way to spend 4 million dollars.

How do you get $4M? 1400 years is 1400*365*24 hours, so you'd use 1400*365*24*60/1000 kWh. A kWh costs around 15 cents (depends where you live). I get about $110k.

I agree that the $4M estimate above is a bit exaggerated. But he did mention "resources" which includes precious human-hours of powerful programmers who volunteered to provide patches for SF.

[jp]

Let us estimate the inveatment in SF project from fishtest data alone from 5 years ago. More than 1 billion games have been tested, 30,000+ patches, and

1400 years of CPU years of computing power have been used.( It means if you run 1 ghz cpu for 1 core computer, you will need 1400 years to complete these tests).

Yes, a long way to spend 4 million dollars.

How do you get $4M? 1400 years is 1400*365*24 hours, so you'd use 1400*365*24*60/1000 kWh. A kWh costs around 15 cents (depends where you live). I get about $110k.

I agree that the $4M estimate above is a bit exaggerated. But he did mention "resources" which includes precious human-hours of powerful programmers who volunteered to provide patches for SF.

Let's not worry about the $ estimate. Is the 1400 CPU years a fair estimate??

Then I'd like to know how many CPU years Lc0 testing has had so far.
The main server shows 115 million games.
Games have also been played on the testserver. The current test run is over 7 million games. I don't know how many previous test runs have been run in parallel with the main server.

Does anyone know the time control or nodes/sec for the training games??

That must all come to way more than 1400 CPU years (1 GHz, 1 core), but what is it roughly?

[grahamj]

Let's not worry about the $ estimate. Is the 1400 CPU years a fair estimate??

Then I'd like to know how many CPU years Lc0 testing has had so far.
The main server shows 115 million games.
Games have also been played on the testserver. The current test run is over 7 million games. I don't know how many previous test runs have been run in parallel with the main server.

Does anyone know the time control or nodes/sec for the training games??

That must all come to way more than 1400 CPU years (1 GHz, 1 core), but what is it roughly?

The LC0 training games use 800 visits (~=nodes) per move. Time depends hugely on hardware, and you really need a GPU, which nearly all of the testers have. My GTX 1070 does around 4000 games/day, or 80 years to do 120M games. In terms of hardware cost and electricity, a GTX 1070 is similar to 6 CPU cores at 4 GHz, so you could say 80*24 ~= 2000 years on something that costs the same as a 1 CPU core at 1 GHz.