THX!

@AOC et al - Do you know of any sort of one-stop book or site that would go through the history, design, types, and best uses of chess engines? Not on the programmer or engineer level perhaps, but something that gives an overview of the above?

If so, TIA!

I read the article too. To me, it seems to say that "breadth" of search is more productive than time-to-depth. But I noticed in their example that it was a clear and exact move choice which had only one "correct" answer. But couldn't that be compared to a puzzle which also has only one correct answer, but is not necessarily a model of real play? IOW, in trying to analyze a string of "moves of interest" (which may or may not have a direct tactical solution) is it possible that time-to-depth might be more useful than breadth of search? 

If you're searching for Captain Kidd's treasure you can dig many shallow holes and cover more ground. But what if the treasure lies just a little deeper? Then you could dig all the holes you want yet never find it.

I just played with Chessify's 300mnps service a bit, and compared it to both four cores of my 3700x and 16 cores of my 5950x. In both cases time to depth was better with my local machines.

I know Chessify has tried to refute this or downplay it previously when I reported it (see link below) but it's still pretty strange. I'm guessing the many cpus needed to get such high nps are running at relatively low speeds, etc. I also understand that SF might search more broadly with more cores at its disposal.

Still, pretty odd.

https://chessify.me/blog/nps-vs-time-to-depth-what-you-should-look-at-when-analy...

Yes, it's time. When you are in an OTB tournament and need to prepare for your next game, sometimes you only have half an hour. I'm not a correspondence player, but I guess depending on the time control they might have more than one game to look at each day.

cathexis wrote on 01/27/22 at 13:54:56:

Letting the engine chug on a single position is not an efficient way to get "the answer". Usually you are better off picking the "best move", or some other move which your human judgment identifies as a "move of interest", playing that move, and then letting the engine calculate. Actually you would play a bunch of moves from the principal variation, let it think a bit, then step backwards until you reach the first position, and see if the engine still ranks the moves the same. In this type of interactive analysis, speed is important, but so is memory, because with a larger cache the engine "remembers" more variations as you change the position slightly.

As a practical matter, it's an arms race. If your opponents don't use an engine at all, you can get by with any old engine on any old hardware. If they use an engine but don't use it well, you can use a budget configuration and still count on out-analysing them. (That's my strategy.) But the stronger you get, the more likely your opponents both have invested in good equipment _and_ know how to use it. What you definitely don't want to do is hammer away on a budget computer, because it probably has budget cooling as well, and a chess engine will peg your CPU at 100%. With electronics, heat is enemy number one.

TonyRo wrote on 01/26/22 at 17:12:00:

... until you factor in *all* the costs of a do-it-yourself solution. Most bean counters (I have that tendency myself) only look at the hardware costs, as pointed out at this link: https://rescale.com/the-real-cost-of-high-performance-computing/

Here's a thought experiment. Suppose you had a single customer who wanted to rent engine analysis time from you. All you have to do is buy a high-end desktop, follow the steps given by Jesse Gersenson in the Laptop for Chessbase & Rybka? thread, and then charge your customer for the engine time. How much would you charge that customer? Don't forget electricity, customer support, OS patches and reboots, etc. Now suppose that customer is you. On the other hand if your business is computer-related, say you already run one or more servers, then the incremental cost of running an "own-cloud" engine is much less. In that case it might be justified.

Chessify is using Oracle servers, and I estimate they are marking up 100% on Oracle's on-demand pricing. But Chessify handles spinning up the engine, gives roughly 24x7 availability, keeps track of time and billing, etc. Basically you just pay and connect. Anyway Oracle sales thinks hundreds of cores is "small", if you wanted to run only one engine they would probably just ignore you.

ChessBase must be using their own servers. If you are using Windows/ChessBase it's quite a reasonable option. One cool thing you can do through playchess.com is set up one machine of your own as a private server, and then access it through another machine of your own. The same as the DIY steps, but playchess.com handles all the details. Requires Fritz or ChessBase on both sides though, which means your "server" actually has to be a Windows desktop. Their other idea of renting out your engine time to others in return for ducats seems a bit gimmicky. I can't imagine how that would be viable long-term.
ChessBase instructions => http://engine-cloud.com/

Nowadays I do think it's worth considering just getting a cheap laptop or desktop (whichever is more appropriate for you), and then just signing up for something like chessify. Their cheapest plan is $30/mo, which seems like a lot, but would cover many years if you are looking to spend big bucks on a chess machine. 

I spent something like $3-4k on a computer years ago (16 cores, dual Xeon CPUs) and get something like 13,000 kN/s with Stockfish 14 NNUE (as a comparison, my friends cheap desktop gets around 2,000-3,000 kN/s). I was thinking about building a beast of a computer using the new AMD chips (32 or 64 cores), but decided to use a Chessify 20% off coupon code and test it for a month. Depending on their server load (presumably), I get around 20,000 to 70,000 kN/s (!) when analyzing, and if I want something other than Stockfish I can throw a bit of extra money at them and use LeelaZero or something weirder like SugaR, etc. 

Something to consider anyway.

The specs is the easy part, just dial everything up to 11. The budget is the hard part.