Stockfish-1.6.2 Benchmarks for 1 to 8 Threads

[mcostalba]

Is this better?

Hi Louis,

I would need the disassembled intermixed with source becasue your is an optimized compile and is very difficult to follow the assembly.

I have sent to you a pm....

Thanks
Marco

[Martin Thoresen]

Hello, I just want to throw in my benchmarks with 1-4 threads that is relevant for this thread.

This is all done with the default settings of:
Minimum Split Depth: 4
Maximum Number of Threads per Split Point: 5

For the hardware used, see this picture:
http://home.halden.net/mordor/3808.jpg

Here I am running the JA official build of 1.6.2 64-bit.

Code: Select all

------------------------------------ 
stockfish-1.6.2 bench 256 t 16 default depth 

for t = 1, 2, 3, 4

1 Thread 
Total time (ms) : 74931 
Nodes searched  : 90964503 
Nodes/second    : 1213976

2 Threads 
Total time (ms) : 50535 
Nodes searched  : 116729314
Nodes/second    : 2309870 

3 Threads 
Total time (ms) : 28074 
Nodes searched  : 93323977 
Nodes/second    : 3324213 

4 Threads 
Total time (ms) : 24008 
Nodes searched  : 101325400 
Nodes/second    : 4220484

As far as my limited math-skills go....

Speedup:
2 threads: 1.9027
3 threads: 2.7382
4 threads: 3.4765

Efficiency:
2 threads: 95.13%
3 threads: 91.27%
4 threads: 86.91%

Is this really that bad? Is for instance Rybka 3 any better?

Best Regards,
Martin

[Martin Thoresen]

To contribute to my above post, here are a few test positions with different settings for:

Minimum Split Depth and Maximum Number of Threads per Split Point.

Giving the engine about 1 minute on the positions.

The following is the end position between Carlsen - Kramnik (1-0) in London Chess Classic 2009.
[d]

Code: Select all

Threads = 4, Hash = 1024
1 min run

Table giving Nodes/Second as a function of Minimum Split Depth and Maximum Number of Threads per Split Point

                                      Max Num Threads per Split Point
			   
	          		            4            5            6            7            8 
			      
Minimum Split Depth = 4    8.790.749    8.844.459    8.804.752    8.847.073    8.756.140

Minimum Split Depth = 5    8.465.013    8.848.837    8.794.021    8.888.137    8.748.415

Minimum Split Depth = 6    8.726.416    8.898.119    8.903.514    8.863.617    8.858.091

Minimum Split Depth = 7    8.525.635    8.840.647    8.790.141    8.790.827    8.768.581

The following the position before white's 20th move between Carlsen - Topalov (1-0) in Nanjing 2009.
[d]

Code: Select all

Threads = 4, Hash = 1024
1 min run

Table giving Nodes/Second as a function of Minimum Split Depth and Maximum Number of Threads per Split Point

                                      Max Num Threads per Split Point
			   
	          		            4            5            6            7            8 
			      
Minimum Split Depth = 4    6.945.467    7.086.768    6.899.610    6.872.362    6.848.492

Minimum Split Depth = 5    6.934.923    6.954.153    7.188.230    6.782.877    6.944.215

Minimum Split Depth = 6    6.910.516    6.975.035    7.024.327    6.851.866    6.938.362

Minimum Split Depth = 7    6.988.451    7.141.546    7.079.926    6.929.858    7.044.496

* didn't find g5

Is any of this conclusive at all? Seems pretty "random" to me.