what kind of operations do you test

For integer testing I do:
- a generic integer manipulation which includes logical and arithmetical operations; +, -, xor, shifts, division and multiplication.
- a quicksort implementation

For floating point:
- floating point Mandlebrot
- floating point vector matrix multiplication (specifically vector rotation with a matrix)

For memory:
- move data using various widths: 8bit, 16bit, 32bit and 64bit

For multiprocessor:
- large matrix by matrix multiplication separated in threads, such that each thread does equal amount of work

All this is written very portable.

The scalability has 2 facets

1) It is designed to take decent amount of time whether you are running it on a 4.77Mhz 8088 or a 4Ghz Core i7 or 1.4Ghz Raspbery Pi.
2) It scales the results back such that the final value returned is comparable between ALL CPUs

Negatives:

1) Code is fairly small, performance will be impacted by good iCache in recent processors
2) I need more tests in each category
3) I need more binaries precompiled for more architectures.

That is why I am trying to open it so that people can run it and I will do that as soon as my company allows me to release open-source.

Here is an example of me running it on a VARIETY of architectures:

For the "(to run)" I actually have numbers somewhere, but they are with an older version of the benchmark so on those machines I need to re-run.
Score - overall score
Int - integer operations score
MP - score at multiprocessor test (this is NOT Int*num_cores, as the tests are different). Only useful to compare MP against various machines.

1                                     OS				Score	Int		MP		Notes
2
3386DX @ 40Mhz							   Linux		   3.12
4Pentium MMX @ 266Mhz					  DOS			  (to run)				
5Pentium II @ 400Mhz						DOS			  (to run)				
6Pentium III @ 833Mhz				  	WinMe		   (to run)
7PowerPC G3 @ 300Mhz						OSX 10.3		 (to run)	
8PowerPC G3 @ 400Mhz						OSX 10.4		 (to run)	
9PowerPC G4 @ 766Mhz						OSX 10.4		 366	 452	  91    1 core
10C2D P8600 @ 2.4Ghz						 OSX 10.8		 (to run)	  
11AMD Phenom(tm) II X4 960T @ 3.0Ghz	Linux 			1768	1834	3532	4 cores
12AMD Phenom(tm) II X6 1100T @ 3.3Ghz  Win10	 		2056	1868	4574	6 cores
13i7-4870HQ @ 2.50GHz				      OSX 10.13		3441	3333	3993	8 cores
14i7-2600 @ 3.40GHz					  	OSX 10.13		3706	3594	4382	8 cores
15i7-4771 @ 3.50GHz						  OSX 10.13		3955	3748	4992	8 cores
16Xeon(R) E5-1650 v2 @ 3.50GHz			OSX 10.13		4363	3814	7235	12 cores
17i7-7700K @ 4.20GHz					    OSX 10.13		4418	4404	6218	8 cores

I am planning to make retrobench public so that people can

a) running it on various architectures
b) compile it for various architectures
c) add more tests

The repository right now provides 7 binaries for different OS/CPU combo: 386, PowerPC G4, PowerPC G3, x64 etc. For OS I currently support: DOS (16bit), Linux (32bit and 64bit), OSX (PowerPC and Intel), Windows (32bit and 64bit). I am working on adding Motorola support.

Does it also work in DOSBOX? And does it work quite accurate in DOSBOX? For example, let say I use your benchmark in a real 386, then I get a certain benchmark score (say, 5762). Then I use your benchmark in DOSBOX while playing with CPU cycles to approach the real 386 score previously obtained using your benchmark (say, 5820).

Would gaming performance between the said 386 and the cycles-set DOSBOX be the same?

Kreshna Aryaguna Nurzaman wrote:

Does it also work in DOSBOX? And does it work quite accurate in DOSBOX? For example, let say I use your benchmark in a real 386, then I get a certain benchmark score (say, 5762). Then I use your benchmark in DOSBOX while playing with CPU cycles to approach the real 386 score previously obtained using your benchmark (say, 5820).

Would gaming performance between the said 386 and the cycles-set DOSBOX be the same?

I suspect it will to some extent. For example I've ran it on my cycle based emulator emulating a 386DX @20Mhz it got me half the speed of my physical 386DX@40Mhz (as expected).

I will run it in DOSBOX as well and see what I get.

For DOS I only have a 16bit executable. I need to get a compiler that produces 32bit executable, as that binary for DOS will be more accurate for anything 386 to Pentium III (in DOS mode).