Elo difference for 4-8-16-24-32 cores

[Sedat Canbaz]

I could be wrong, but I don't think blitz is likely to bring out the full advantage of lots of cores. You're more likely to see it with a longer time control. But as always, we need to make compromises as we don't have unlimited resources.

Hello Ray,

No...it depends and I suggest to see my Blitz results

And as we see...the Top MP engines need updates, optimizations for the latest new fast decent machines (for 12 cores, 16 cores; 24 cores etc...)

And here is another SCCT Auto232 comparison, so up to 6 cores or maybe up to 8 cores is ok...
But later....with more cores, the MP test results confirm differently (not so good)

Code: Select all

Rank  Name                   Elo    +    -  games  score oppo. draws
Houdini 2.0c Pro x64 6c      3423   19   18  1008   70%  3261   42%
Houdini 2.0 Pro x64 4c       3363   16   16  1131   61%  3294   49%
Houdini 2.0c Pro x64 1c      3263   13   13  2088   61%  3186   42% 

*Houdini's Elo Difference (between 1 core vs 6 cores) = 160 Elo

Code: Select all

Rank  Name                    Elo    +    -  games  score oppo. draws
Rybka 4.1 x64 6c             3358   15   15  1363   59%  3305   54%
Rybka 4.1 x64 4c             3293   13   14  1602   47%  3314   56%
Rybka 4.1 x64 1c             3199   13   13  1745   48%  3213   49% 

*Rybka's Elo Difference (between 1 core vs 6 cores) = 159 Elo

Note: Auto232 matches with Ponder ON, same short neutral book; TC: 4min+2sec etc...


For More Details:
https://sites.google.com/site/computers ... ct-auto232



Greetings,
Sedat

[Sedat Canbaz]

More comparisons by SCCT 4m+2s, CCRL 40/4 and CEGT 40/4

CCRL
Houdini 2.0c 64-bit 4CPU 3273
Houdini 2.0c 64-bit 1CPU 3197

CEGT
Houdini 2.0c x64 4CPU 3156
Houdini 2.0c x64 1CPU 3052


CCRL:Houdini's Elo Difference (between 1 core vs 4 cores) = 76 Elo
CEGT:Houdini's Elo Difference (between 1 core vs 4 cores) = 104 Elo
SCCT:Houdini's Elo Difference (between 1 core vs 4 cores) = 100 Elo

Note: again there is something wrong with CCRL !!)
Sorry to say that but reality is reality...

CCRL
Rybka 4.1 64-bit 4CPU 3198
Rybka 4.1 64-bit 1CPU 3100

CEGT
Rybka 4.0 x64 4CPU 3071
Rybka 4.0 x64 1CPU 2984

CCRL:Rybka's Elo Difference (between 1 core vs 4 cores) = 98 Elo
CEGT:Rybka's Elo Difference (between 1 core vs 4 cores) = 87 Elo
SCCT:Rybka's Elo Difference (between 1 core vs 4 cores) = 94 Elo

Rybka (with all 3 rating lists) seems to be ok...!


Hopes helps...

Best,
Sedat

[Leto]

I could be wrong, but I don't think blitz is likely to bring out the full advantage of lots of cores. You're more likely to see it with a longer time control. But as always, we need to make compromises as we don't have unlimited resources.

Hello Ray,

No...it depends and I suggest to see my Blitz results

And as we see...the Top MP engines need updates, optimizations for the latest new fast decent machines (for 12 cores, 16 cores; 24 cores etc...)

And here is another SCCT Auto232 comparison, so up to 6 cores or maybe up to 8 cores is ok...
But later....with more cores, the MP test results confirm differently (not so good)

Code: Select all

Rank  Name                   Elo    +    -  games  score oppo. draws
Houdini 2.0c Pro x64 6c      3423   19   18  1008   70%  3261   42%
Houdini 2.0 Pro x64 4c       3363   16   16  1131   61%  3294   49%
Houdini 2.0c Pro x64 1c      3263   13   13  2088   61%  3186   42% 

*Houdini's Elo Difference (between 1 core vs 6 cores) = 160 Elo

Code: Select all

Rank  Name                    Elo    +    -  games  score oppo. draws
Rybka 4.1 x64 6c             3358   15   15  1363   59%  3305   54%
Rybka 4.1 x64 4c             3293   13   14  1602   47%  3314   56%
Rybka 4.1 x64 1c             3199   13   13  1745   48%  3213   49% 

*Rybka's Elo Difference (between 1 core vs 6 cores) = 159 Elo

Note: Auto232 matches with Ponder ON, same short neutral book; TC: 4min+2sec etc...


For More Details:
https://sites.google.com/site/computers ... ct-auto232



Greetings,
Sedat

Thanks Sedat. Your test is showing a 60 elo improvement for Houdini 2.0c going from 4 cores to 6 cores. Why would my test of Houdini 4 show barely any difference going from 4 to 8 cores?

Here's some possibilities:
1. Maybe going from 6 to 8 cores is hurting Houdini 4
2. Perhaps the SMP implementation in Houdini 4 is worse than in Houdini 2?
3. Perhaps my settings are wrong? My systems are Dual Xeon x5650 with HT off (12 cores), Dual Xeon E5620 with HT off (8 cores), and an AMD 8300 (8 cores). For CEGT testing I use default settings, which for Houdini 4 is Numa on, Numa Offset 0, Split Depth 10.

[Sedat Canbaz]

Thanks Sedat. Your test is showing a 60 elo improvement for Houdini 2.0c going from 4 cores to 6 cores. Why would my test of Houdini 4 show barely any difference going from 4 to 8 cores?

Here's some possibilities:
1. Maybe going from 6 to 8 cores is hurting Houdini 4
2. Perhaps the SMP implementation in Houdini 4 is worse than in Houdini 2?
3. Perhaps my settings are wrong? My systems are Dual Xeon x5650 with HT off (12 cores), Dual Xeon E5620 with HT off (8 cores), and an AMD 8300 (8 cores). For CEGT testing I use default settings, which for Houdini 4 is Numa on, Numa Offset 0, Split Depth 10.

Not at all Leto


Just to make it more clear,
- Hoodini 2.0c 1 core and Rybka 4.1 1 core are played on my i7 920 and QX9650 machines
- Also Houdini 2.0c 4 cores and Rybka 4.1 4 cores are played on my i7 920 and QX9650 machines
- The chess speed of the both hardwares (i7 920@3.0GHz/QX9650@3.66GHz) are almost identical
- Houdini 2.0c 6 cores and Rybka 4.1 6 cores are played on my i7 980X and i7 970 machines
* MP matches on same PC and Ponder OFF were not allowed...only in Auto232 mode, !


For examples, see please Fritz Chess Benchmarks:

Code: Select all

Hardware-Processor        Speed      Cores     kN/s
Intel Core i7 970       @ 4.00 GHz     6      17281
Intel Core i7 920       @ 3.30 GHz     4      10350
Intel Core 2 QX9650     @ 3.66 GHz     4      10253

Yes....the Elo difference (between i7 920@3.0GHz vs i7 970 @4.0GHz) is approx. 60 Elo
Another note: the speed difference (between i7 920@3.0GHz vs i7 970 @4.0GHz) is approx. 1.64
So... under these conditions, in case of doubling the speed is expecting at least 70-80 Elo

And there is one true,
i7 6-core machines are still beast for chess!

Hopes this helps too


About rest issues...
MP engine creators views and statements will be more more useful...


Best,
Sedat

[Sedat Canbaz]

Dear Leto,

Btw,
We have no any Auto232 testings about the Elo difference between AMD 8300 8 cores vs Intel i7 6 cores

There is a possibility,
Houdini MP or any other MP engine suffers on AMD 8300 8 cores machines

And now I remembered (I read somewhere over forums) that Amd's cores are not identical as Intel ones....!
Actually even the chess benchmarks show that that Intel's 6 cores are faster and better...

So my view for AMD 6 cores or 8 cores are in chess strength minus plus as Intel Quads !



Best,
Sedat

[M ANSARI]

That was quite some time ago! I gave that computer to my brother as he needed something fast for his video editing. I would like to build something new and powerful for chess but have not really been following up on what is the latest hardware. I still have that powerful dual socket vapor cooling system and I think with slight modifications I could easily make it use the latest sockets. Is there are over clockable dual socket system? Would be nice to have something running 2 x 8 cores at 5 Ghz or more !!! I did have my old dual socket system running @ 4.8 Ghz on 8 cores for a long time and I had thought I sent the benchmark using that. At the moment it is running @ 4 Ghz on water though as the compressors of vapor is just too noisy for general home use. What is amazing is how stable that system was under heavy 24/7 use even when overclocked. I think the native speed was 3.2 Ghz but I never ran it at less than 4 Ghz ... and it has been almost 4 or 5 years now!!!

Here is a picture of the vapor cooling unit ... nice chilly -56C per processor

[Leto]

Dear Leto,

Btw,
We have no any Auto232 testings about the Elo difference between AMD 8300 8 cores vs Intel i7 6 cores

There is a possibility,
Houdini MP or any other MP engine suffers on AMD 8300 8 cores machines

And now I remembered (I read somewhere over forums) that Amd's cores are not identical as Intel ones....!
Actually even the chess benchmarks show that that Intel's 6 cores are faster and better...

So my view for AMD 6 cores or 8 cores are in chess strength minus plus as Intel Quads !



Best,
Sedat

No I don't think it's the AMD system causing the issue, with Houdini 4 on my two Intel Dual Xeon machines I'm still getting barely any performance increase going from 4 cores to 8 or 12 cores. And other engines such as Stockfish and Komodo do benefit going from 4 cores to 8 on my AMD system, so it's definitely not that.

I do agree with you that the Intel systems are superior, but that shouldn't make a noticeable difference in my testing because I don't do multiple pc testing.

Maybe it's the settings. What is Numa? Should I leave it enabled for Houdini 4 on my three systems?

[Sedat Canbaz]

That was quite some time ago! I gave that computer to my brother as he needed something fast for his video editing. I would like to build something new and powerful for chess but have not really been following up on what is the latest hardware. I still have that powerful dual socket vapor cooling system and I think with slight modifications I could easily make it use the latest sockets. Is there are over clockable dual socket system? Would be nice to have something running 2 x 8 cores at 5 Ghz or more !!! I did have my old dual socket system running @ 4.8 Ghz on 8 cores for a long time and I had thought I sent the benchmark using that. At the moment it is running @ 4 Ghz on water though as the compressors of vapor is just too noisy for general home use. What is amazing is how stable that system was under heavy 24/7 use even when overclocked. I think the native speed was 3.2 Ghz but I never ran it at less than 4 Ghz ... and it has been almost 4 or 5 years now!!!

Here is a picture of the vapor cooling unit ... nice chilly -56C per processor

Yes...once more I remembered again the old good days!)

Wow....-56C per processor !!

It looks like Siberia ))


Best,
Sedat

[Sedat Canbaz]

No I don't think it's the AMD system causing the issue, with Houdini 4 on my two Intel Dual Xeon machines I'm still getting barely any performance increase going from 4 cores to 8 or 12 cores. And other engines such as Stockfish and Komodo do benefit going from 4 cores to 8 on my AMD system, so it's definitely not that.

I do agree with you that the Intel systems are superior, but that shouldn't make a noticeable difference in my testing because I don't do multiple pc testing.

Maybe it's the settings. What is Numa? Should I leave it enabled for Houdini 4 on my three systems?


NUMA support (Houdini Pro)
Top Previous Next

Most CPU mother boards with multiple sockets employ the so-called "NUMA" architecture.

Houdini Pro detects the NUMA configuration at start-up and will adapt its memory management and thread interaction based on the different NUMA nodes that are available.

As I mentioned before,
It seems Houdini suffers on your AMD machine...
I think Robert Houdart can inform us more...
And if Houdini suffers....there is a possibility other MP chess engines to suffer too

Btw, here is some info about Numa (from Houdini's site):
http://www.cruxis.com/chess/manual/inde ... upport.htm


NUMA support (Houdini Pro)

Most CPU mother boards with multiple sockets employ the so-called "NUMA" architecture.

Houdini Pro detects the NUMA configuration at start-up and will adapt its memory management and thread interaction based on the different NUMA nodes that are available.

Numa (checkbox)

Enable or disable the NUMA-awareness. (Enabled by default)

When NUMA is enabled, Houdini will organize memory and threads to take into account the NUMA configuration of the hardware. This can significantly enhance the scaling of the engine beyond 6 cores.

Combined with Large Pages the speed gain can be up to 25% depending on the number of cores, the motherboard and CPU brand.

See the Houdini Pro topic for some real performance data obtained on a 16-core dual AMD Opteron-6128 box.

Numa Offset

Default 0, min 0, max 16.

The NUMA offset is useful if you're running multiple instances of Houdini, in which case you should assign a different NUMA node for each running Houdini process to avoid the Houdini instances competing for the same resources.

For example, on a 16-core hardware with 4 NUMA nodes, you could be running four Houdini processes each using 4 cores. By setting Numa Offset to 0, 1, 2 and 3, each Houdini instance will be running on its own NUMA node without conflicting with the other Houdini instances.

[Modern Times]

No I don't think it's the AMD system causing the issue, with Houdini 4 on my two Intel Dual Xeon machines I'm still getting barely any performance increase going from 4 cores to 8 or 12 cores. And other engines such as Stockfish and Komodo do benefit going from 4 cores to 8 on my AMD system, so it's definitely not that.

Agreed - AMD won't be the issue. Robert used to do all his development and testing on AMD in fact.

Larry K also believes that AMD actually favours Houdini vs Komodo.