For how long will Stockfish be competitive versus the best NN ?

[jp]

That was kind of my point - for average sort of money that an average user would spend, Stockfish may well perform better than an NN engine. Getting the best out of NN engines is a very expensive endeavour. The £600 price point was pretty low but for a little more maybe £800 you'd get a machine with something with a GTX 1660 plus perhaps. Anyway, I'm not in the market for a new PC myself, and when I will be it will be a laptop at £600 - £800 current prices. That will just be an integrated GPU.

For a "balanced" mass-market budget PC, the estimated "fair" ratio of (nps Stockfish)/(nps Leela) is:

~8400 (based on budget non-gaming PC).

[See http://talkchess.com/forum3/viewtopic.p ... 18#p828418 and following posts.]

For typical testing of NN engines vs SF, where they are competitive, the ratio used is only ~875,

i.e. they give the NN engines about 10 times "too much" computer power relative to budget machines, so for sure SF will be stronger on the budget machines.

[Chessqueen]

That was kind of my point - for average sort of money that an average user would spend, Stockfish may well perform better than an NN engine. Getting the best out of NN engines is a very expensive endeavour. The £600 price point was pretty low but for a little more maybe £800 you'd get a machine with something with a GTX 1660 plus perhaps. Anyway, I'm not in the market for a new PC myself, and when I will be it will be a laptop at £600 - £800 current prices. That will just be an integrated GPU.

For a "balanced" mass-market budget PC, the estimated "fair" ratio of (nps Stockfish)/(nps Leela) is:

~8400 (based on budget non-gaming PC).

[See http://talkchess.com/forum3/viewtopic.p ... 18#p828418 and following posts.]

For typical testing of NN engines vs SF, where they are competitive, the ratio used is only ~875,

i.e. they give the NN engines about 10 times "too much" computer power relative to budget machines, so for sure SF will be stronger on the budget machines.

WOW 2085 people viewing this : https://tcec-chess.com/

[Ovyron]

Whatever you spend you can still get great analysis by simply letting the machine think longer. Now, it might be a very long wait if you want the best answer in the world, but you spend what you feel comfortable and live with that.

For people with this mindset, let me ask you, after leaving the computer analyzing for a very long time, and getting a great output out of it: now that you have the output, was there a way for you to find this output any other way? Was it possible to get it in a shorter time frame? Was it possible to find it in a MUCH shorter time frame?

And the answers are always "yes, yes, and yes." People don't need better hardware because if they learned how to interact with the positions they could get to that "best answer in the world" with what they already have. If in school kids were taught how to interact with chess engines to find the best answers with slow hardware, most people would know how to do it, and it wouldn't be deemed harder than... algebra.

But since this is something that nobody teaches, and there's no tutorial on how to learn it, people live with the impression that they're things you can't do unless you have big hardware. Fastest mates were put as an example of something that requires big hardware, but I'm talking about something useful, like playing the best move on a position, that doesn't need it.

[Chessqueen]

That was kind of my point - for average sort of money that an average user would spend, Stockfish may well perform better than an NN engine. Getting the best out of NN engines is a very expensive endeavour. The £600 price point was pretty low but for a little more maybe £800 you'd get a machine with something with a GTX 1660 plus perhaps. Anyway, I'm not in the market for a new PC myself, and when I will be it will be a laptop at £600 - £800 current prices. That will just be an integrated GPU.

For a "balanced" mass-market budget PC, the estimated "fair" ratio of (nps Stockfish)/(nps Leela) is:

~8400 (based on budget non-gaming PC).

[See http://talkchess.com/forum3/viewtopic.p ... 18#p828418 and following posts.]

For typical testing of NN engines vs SF, where they are competitive, the ratio used is only ~875,

i.e. they give the NN engines about 10 times "too much" computer power relative to budget machines, so for sure SF will be stronger on the budget machines.

WOW 2085 people viewing this : https://tcec-chess.com/

Are you saying that the NN has a power advantage over Stockfish? Not if you count the Speed or Depth/SD, or even Nodes; compare it take a look at it closely https://tcec-chess.com/

[jp]

Are you saying that the NN has a power advantage over Stockfish? Not if you count the Speed or Depth/SD, or even Nodes; compare it take a look at it closely https://tcec-chess.com/

The NN engine in all these contests has a massive computer hardware advantage over Stockfish by any of the metrics that are discussed in the thread that is linked (i.e. money, electricity consumption, "balanced" off-the-shelf computers, same CPU with CPU version of Leela, etc.). The "nodes" count is meaningless. The definition of "node" is very different for NN engines (but even if it were the same it would be meaningless). They have conveniently defined it in a way that vastly undercounts (compared with AB engines). This is all discussed in the linked thread. Just look at that thread.

[Milos]

GPUs can also be used for other interesting things like folding proteins, searching for Mersenne Primes, astronomical or physics grid calculations, etc.

Or you can play computer games that is like million times more interesting than folding proteins or what not .

For instance, I bought an AMD 3970x which is 1/3 slower than an AMD 3990x. So if I analyze one and a half times as long, I have the same approximate output. Whatever you spend you can still get great analysis by simply letting the machine think longer. Now, it might be a very long wait if you want the best answer in the world, but you spend what you feel comfortable and live with that.

3970x is atm the best price/performance non-budget config. You can assemble the full system for like $800 (with 32GB of RAM). If you want budget config you take 2700x with 16GB RAM (at least in Europe) for $500.

On an inexpensive smartphone, you can get stockfish and have a GM in your pocket to analyze positions for you at your beck and call. How much more power you need is strictly up to you. If you just want a critique of your game, you don't need a $5,000 system, since your cell phone will do nicely.

A slow 8 core smartphone gives like 3-4M nps (SF11) from starting position. That is like 2 fast CPU cores. So at worst only 150Elo below your 3970x machine. That is not GM level, that is 600 Elo above GM level .

[Dann Corbit]

A slow 8 core smartphone gives like 3-4M nps (SF11) from starting position. That is like 2 fast CPU cores. So at worst only 150Elo below your 3970x machine. That is not GM level, that is 600 Elo above GM level .

Amen, brother.
A 1944 v2 rocket arrived at 4000 MPH. Not much improvement since then. That's a long time.

[Uri Blass]

Whatever you spend you can still get great analysis by simply letting the machine think longer. Now, it might be a very long wait if you want the best answer in the world, but you spend what you feel comfortable and live with that.

For people with this mindset, let me ask you, after leaving the computer analyzing for a very long time, and getting a great output out of it: now that you have the output, was there a way for you to find this output any other way? Was it possible to get it in a shorter time frame? Was it possible to find it in a MUCH shorter time frame?

And the answers are always "yes, yes, and yes." People don't need better hardware because if they learned how to interact with the positions they could get to that "best answer in the world" with what they already have. If in school kids were taught how to interact with chess engines to find the best answers with slow hardware, most people would know how to do it, and it wouldn't be deemed harder than... algebra.

But since this is something that nobody teaches, and there's no tutorial on how to learn it, people live with the impression that they're things you can't do unless you have big hardware. Fastest mates were put as an example of something that requires big hardware, but I'm talking about something useful, like playing the best move on a position, that doesn't need it.

I totally disagree.
The answer is always no.

1)Fastest mate is clearly useful and there may be positions when the not playing the fastest mate is a draw by the 50 move rule.
2)I am sure that there are positions that you will be slower in finding the best move by your methods
otherwise you can get a huge software improvement by simply writing an engine that find the best moves faster.

It was probably possible many years ago when the search algorithm of engines was stupid but not today.

[Morten Lohne]

You can't just put a top end GPU card into a very inexpensive PC because often the power supply is not up to the task.

Just add a slightly more powerful PSU then, they're not expensive.

However, you can get a much cheaper card that uses less power and still provides very good analysis.
https://www.videocardbenchmark.net/common_gpus.html
Look at the price performance graph and you can see for $200 you can get a pretty good card for LC0 that will also use a lot less power than a 2080Ti.

While this is true for gaming, it is simple not true for running Leela. GPUs with special NN inference hardware will crush any GPU without without it. For now, your only choices for that are Nvidia RTX cards (2060, 2070, 2080 and their cousins). For instance, an RTX 2060 gets more than 3x more nps than a GTX 1070 ti, even though they are roughly equivalent for gaming.

[Dann Corbit]

You can't just put a top end GPU card into a very inexpensive PC because often the power supply is not up to the task.

Just add a slightly more powerful PSU then, they're not expensive.

However, you can get a much cheaper card that uses less power and still provides very good analysis.
https://www.videocardbenchmark.net/common_gpus.html
Look at the price performance graph and you can see for $200 you can get a pretty good card for LC0 that will also use a lot less power than a 2080Ti.

While this is true for gaming, it is simple not true for running Leela. GPUs with special NN inference hardware will crush any GPU without without it. For now, your only choices for that are Nvidia RTX cards (2060, 2070, 2080 and their cousins). For instance, an RTX 2060 gets more than 3x more nps than a GTX 1070 ti, even though they are roughly equivalent for gaming.

The 1650 super is about half as fast as a 2080 super, and it is less than $200:
https://www.videocardbenchmark.net/comp ... 4123vs4167
It is very new, so it has a modern AMD instruction set.
Yes, it will not be as fast as a 2080 Super, but if you wait two or three times as long, your answer will be just as good. I have two 2080 Supers in the machine right next to me, but not everyone wants to spend a lot of money in order to push little wooden horsies around on some cardboard squares.