Amazon EC2 cloud

[Peperoni]

An update of that
Which servers would be the best to run games Stockfish vs Stockfish for example?
The C5a look good to me, but I am not an expert ...

[mvanthoor]

no, its not cheap but the easier solution, i dont want buy every 2 years a new PC again and again.

Then get yourself a good case, and power supply, a good mainboard, a powerful 16-core CPU (at least; so you can run 16 games/32 engines at the same time), and 32 - 128 GB of RAM depending on how big you want your hash tables to be (on short time controls, 512 MB per engine is enough). Install Linux on it (Mint if you want to make it easy, Debian Testing if you want to build from scratch), with VNC server and control it from your desktop computer.

More cores is more important than raw speed per core, because with more cores you can run more games.

A computer like that can sit in a corner for 6-10 years before you decide to replace it (the more powerful you make it, the less likely you feel the need to have a more powerful computer to make testing faster).

At least, the above is what I intend to do, as soon as I decide if I'm going to risk my neck with Zen 5000, or wait for Intel to come up with a successor of the 10980XE (or, more precisely, a new chipset for these CPU's).

[jdart]

If you are looking at Amazon instances, what you probably want for chess is the c6g instances (see https://aws.amazon.com/ec2/instance-types/).

But it gets pretty expensive pretty fast running these. They are good for short-term use if you don't want to buy a box.

--Jon

[Peperoni]

If you are looking at Amazon instances, what you probably want for chess is the c6g instances (see https://aws.amazon.com/ec2/instance-types/).

But it gets pretty expensive pretty fast running these. They are good for short-term use if you don't want to buy a box.

--Jon

May I ask why these CPU in particular?

[jdart]

Designed for high compute loads.

[Peperoni]

Designed for high compute loads.

C6g or C6gd?

[jdart]

C6g or C6gd?

I think the only difference is that C6gd instances have local very fast storage (NVMe). Standard AWS storage systems are already quite decently fast though.

[Peperoni]

Yes, I am not sure it makes a difference. Is the disk bandwith used only when the PGN file are being written?
Or does it still have an impact during the computation?

[mvanthoor]

BTW; if one wants to build a system with a lot of cores, a cluster would be possible. I saw a post by the Zurichess author ( https://chess.stackexchange.com/questio ... h-2000-elo ) where he said he built an ODroid cluster. Those boards have 4 cores each, and he uses 6, for 24 cores.

Newer ODroids even have 8 cores, so if you use 5 of them, you could build a 40 core cluster for not too much money, and test 40 games at once.

This would be a lot cheaper than building a 16+ core computer for those tests (let alone a 32+ core one). You can test very fast because you can run 36 games or so at once (leave some cores for the OS and other software), but because the cores are slower than the ones in a PC, the engines will reach less depth; the quality per game will therefore be less.

I'm sorry to say that I cannot determine if the quality will be high enough for testing, but I can verify that Stockfish running on my 3 year old midrange phone (Nokia 6.1) on a single core reaches depth 16 almost instantly.

[jdart]

I run matches distributed over several machines. Kind of like Stockfish does, but mine are all on a LAN. The machines include a server box with 4 independent 32-core nodes, for 128 cores total. Then there's a 32-core and a 24-core, so 184 cores total. The 32-core is new but the others are used, which really lowers the cost.