TCEC resurrection - need to decide between ponder ON or OFF

[Martin Thoresen]

Well, you just pointed out that your goal is entertainment, not benefits. Pondering is even more detrimental than hyper threading. Thie way I proposed you can award engines with (better) SMP implementation.

And yes, it seems something has happened: the latest reports I have seen here claimed that Houdini on i7 would run about 20% faster nps-wise with HT on, which is just enough to compensate for the search overhead from doubling the number of threads. So it seems that HT is about neutral. But who cares whether you run 10% faster or slower? That should have no impact on entertainment value.

Finally, I thought that Polyglot did support affinity? Or has this been deprecated again? Normally, you should be able to instruct the GUI to define engine priorities and affinities. Not set them by hand...

You may be right about HT, I just did a Fritz Chess Benchmark test:
With 4 cores HT off, the result was about 13500 kn/s.
With 8 cores HT on, the result was about 16500 kn/s.

Then perhaps 3 cores ponder ON is also a viable alternative.

[jpqy]

Hi Martin,

Is same what i told you on page 2!

You get +22,22..% higher nodes/sec. -> Try some games

JP.

[Martin Thoresen]

Hi Martin,

Is same what i told you on page 2!

You get +22,22..% higher nodes/sec. -> Try some games

JP.

I am running a test tournament now with 3 cores Ponder ON per engine, with HT ON.

The problem is that currently ChessGUI is having trouble writing the correct values in the live pgn file when using Ponder ON, comparing to looking at the GUI directly...

[Modern Times]

Even though people have 4 cores (or the more "bogus" 6-8 cores from AMD), one should consider the speed throughput. The latest Intel (Ivy Bridge) is still far superior to anything AMD have, clock for clock.

I'd rather have an 8-core AMD piledriver than an Intel Ivybridge any day. There is some shared resource (4 modules, 8 cores) but I'd still call it an 8-core CPU.

[Martin Thoresen]

I'd rather have an 8-core AMD piledriver than an Intel Ivybridge any day. There is some shared resource (4 modules, 8 cores) but I'd still call it an 8-core CPU.

Ray, I made it sound more negative than what was intended.

Except for slighter lower performance, the major problem with AMD these days is the power consumption (in my opinion).

Here's a piledriver fritz chess benchmark:
http://news.softpedia.com/newsImage/AMD ... zer-3.jpg/

And here's a power consumtion chart (compare to my 3770k, for example):
http://www.bit-tech.net/hardware/2012/1 ... 0-review/7

[Lavir]

Why don't you turn hyper-threading on, and set affinity for engine A to cores 1, 3 and 5, and of engine B to cores 2, 4, 6? Then you can run them at 3 cores with ponder on (and still have 1 physical core, (7 and 8) to spare).

In that way one engine will get all physical cores and the other all the logical ones, so actually giving one of the two (the latter) a great advantage. It's true that the tournament will be for entertainment, but surely entertainment also means somewhat same rules for every participant, until you don't want to purposely handicap some engines in confront to others.

A better affinity is 0,1,2 for the first and 4,5,6 for the second. In this way both get 2 physical + 1 logical core to compute on.

[Lavir]

And yes, it seems something has happened: the latest reports I have seen here claimed that Houdini on i7 would run about 20% faster nps-wise with HT on, which is just enough to compensate for the search overhead from doubling the number of threads.

True, but nobody till now has made tests on 2 cores + 1 threads or 1 core + 2 threads or similar things, so it is not known the impact of HT on those cases.

The tests where all done using same amount of cores and threads, not mixed.

[Lavir]

[
Then perhaps 3 cores ponder ON is also a viable alternative.

It can be, but you will have to set affinities manually every time elsewhere one of the two engines will get more physical processes than the other much probably.

As I specified above a good setting would be 0,1,2 for one and 4,5,6 for the other (assuming 4 cores).

Now the question is: do you want to manually adjust affinities every time a match is started?

And also if this is matches will be for entertainment purposes as I said nobody till now tested the difference with mixed cores/threads and the impact it would have on playing strength, so it can also happen that something strange could happen; you must take this in consideration.

[hgm]

True, but nobody till now has made tests on 2 cores + 1 threads or 1 core + 2 threads or similar things, so it is not known the impact of HT on those cases.

The tests where all done using same amount of cores and threads, not mixed.

Well, what I proposed would have each core run 1 thread for 1 engine, and 1 thread for the other engine. So each engine would have three threads (on three different, otherwise fully loaded cores). Not any of the things you mention.

And, like I said, who cares about the impact? As long as you make both engines do the same thing, it will be fair. No matter how much they slow down or speed up by it.

[Lavir]

Well, what I proposed would have each core run 1 thread for 1 engine, and 1 thread for the other engine. So each engine would have three threads (on three different, otherwise fully loaded cores). Not any of the things you mention.

You proposed "engine A cores 1, 3 and 5, and of engine B to cores 2, 4, 6" and in this proposition engine A gets all the logical cores (1,3,5 are all logical cores) while B all the physical ones (2,4,6).

In a 4 CPU cores 0,2,4,6 are the physical ones, 1,3,5,7 the logical. If you want to give only 1 real physical + 2 threads then you must give something like 0,1,3 to one and 2,5,7 to the other. But I think it's better 2 physical + 1 thread.