Mandelbrot Benchmark

[Kuemmel]

Hi there,

after lots of optimization work I finished now an almost final version of my Mandelbrot Benchmark. It focuses on FPU and SSE2 units of current processors, while neglecting almost all influence of cache and memory subsystem and graphics card. I used a lot of time implementing the whole thing completely in assembler and used lots of techniques to make it really fast with loop unrolling, independent instruction chains, multi core support using the benefits of all the available CPU's.

I think it really shows how over the years theses calculation units where improved or not...I have a column on my webpage for efficiency per core per MHz.

Any comments and test results welcome...

@@@ New Version released 01-Jun-2008 @@@

Direct link:
http://www.mikusite.de/x86/KMB_V0.53H-32b-MT.zip
Homepage with results:
http://www.mikusite.de/pages/x86.htm

[3lfk1ng]

Mandelbrots are awesome, I'll have to check this out later.

[MuffinFlavored]

SSE2:
3365.647

FPU:
1095.311

Fix it so that if you minimize the benchmark window, it doesn't come back a big black square that won't go away.

[fellix_bg]

Sse2: 2645.455

Fpu: 843.686

[Metric]

Q6600 (400x9.0)
SSE2: 4711.906
FPU: 1504.404

[s1nykuL]

FPU = 680.446
SSE2 = 2182.936

C2D E6400@8x400=3.2GHz. 1:1 FSB.

Kuemmel, that's a quick algorithm.


ChaosPro is a good fractal proggy to play with.

[Major_A]

My kind of benchmark... quick.

[Kuemmel]

Thanks for all the results !

I'm trying hard to find somebody with a P4 or Xeon based P4 CPU with Hyperthreading...anybody out there ? Or can you pass it to somebody with that kind of CPU ??? Pleeeeeeeaaaaase... ;-)

[3lfk1ng]

Oh, we have a bunch of 3.2ghz P4's with HT in my lab on campus. Ill give it a shot next Tuesday.

[Entity_Razer]

Dual Xeon 2.4Ghz with hyperthreading here

Will Test tomorrow. To tired right now. if I forget just shoot me a PM

[Bobsama]

I was trying to get you results from an old K6 Chomper Ext at 550MHz, but it didn't work.

[Kuemmel]

I was trying to get you results from an old K6 Chomper Ext at 550MHz, but it didn't work.

Hi, I think that's possible and okay, becuase the K6 neither supports the additional FPU instructions back in time invented with the incarnation of "Pentium Pro" and "Pentium II" and of course no SSE2 also at that time...sorry...may be should have implemented a kind of check and warning...

...by the way...you can send just any results, also AMD FX and X2 and Phenoms are very interesting !!!

[Qwik]

cool

[The0men]

cool

Just wondering why you score so low at 4ghz?
Using vista?

Got Q6600 @ 3.645 mem at 996 4-4-4 and score of:

1533.97
4858.58

Very nice benchmark though, love quick ones.

[Kuemmel]

Just wondering why you score so low at 4ghz?
Using vista?...
...Very nice benchmark though, love quick ones.

Hm, yep, your result is like expected, his result is a bit slow...I had one result where Vista 32 bit was slow and Vista 64 bit was okay, but also another Vista 32 bit result where everything was okay...no conclusion yet...may be other programs running or something ?

Because normally the benchmark really scales totally in line with the clock speed of a processor generation, so all Core 2 scale the same.

@EDIT: in other words, at 4050 MHz with a Core 2 Quad I think a result up to 5350-5400 shoule be possible independant of the memory setup of the system or motherboard...

[Qwik]

Just wondering why you score so low at 4ghz?
Using vista?

Got Q6600 @ 3.645 mem at 996 4-4-4 and score of:

1533.97
4858.58

Very nice benchmark though, love quick ones.

Nope I am using XP32 .. and I ran bench @ everyday settings with other programs running like normal(virus scan, skype, EI7, etc..). I'll rerun it
by it self see what a difference it makes.

it did.

[CyberDruid]

Interesting result



I score about 300 points higher if I have all 8 cores at 100% load Folding

Turn off Folding...score drops.

[jspace]

Is there a reason I'm getting TR/Crypt.XPACK.Gen warnings?

[Metric]

Is there a reason I'm getting TR/Crypt.XPACK.Gen warnings?

False positive. From the thread in Benchmarking:...

Yep, known problem, it's a false alarm. I loaded my program up on "www.virustotal.com" and only got 3 kind of positive or warning from 32...it's a common problem within the users of FlatAssembler for coding...plenty false alarams there...

[Lestat]

Q6600 G0 @ 3.835

i wanna see someone do Mandel's that are created or benchmarked as 100% GPU based,, thats G P U
see what the power of the GPU can do vs CPU

[Kuemmel]

Q6600 G0 @ 3.835
i wanna see someone do Mandel's that are created or benchmarked as 100% GPU based,
see what the power of the GPU can do vs CPU

Yep, interesting topic...the problem is that at the moment the floating point unit on a GPU is only single precision and when you go 'deeper' into the Mandelbrot 'land' you need double precision...I read that the GPU producer might change this may be in the future...here's a refernece:

http://www.theinquirer.net/gb/inquir...gpgpu-monsters

...but I don't know what's the actual state...so it's more for the future, once these things are spread...

[Kuemmel]

I score about 300 points higher if I have all 8 cores at 100% load Folding Turn off Folding...score drops.

Hm, what is acutally "load Folding" ? I didn't find anything directly usefull while googleing it...where do you set this thing and is it normally supposed to be turned on for everyday use or something ?

[The0men]

He means when his processors are being used at %100 load whilst he is folding, which is just a program using your CPU to do equations for various tasks like scientific research, or medical research, or biological etc.

Just like running a stress test like prime on all 8 of his cores.

[Oldguy932]

Laptop: 1.73ghz Pentium M

FPU: 171.636
SSE2: 190.112

Desktop: Phenom 9600 (TLB fix disabled) 258x11.5=2967

FPU: 1036.228
SSE2: 2628.747

It seemed kind of short, taking only like 10 seconds to run on the phenom. I'm not completely familiar with mandelbrots, but maybe you could make the graphs large by changing the functions some so the cpu has to calculate alot more. Very nice job so far though.

[Kuemmel]

Desktop: Phenom 9600 (TLB fix disabled) 258x11.5=2967

FPU: 1036.228
SSE2: 2628.747

It seemed kind of short, taking only like 10 seconds to run on the phenom. I'm not completely familiar with mandelbrots, but maybe you could make the graphs large by changing the functions some so the cpu has to calculate alot more. Very nice job so far though.

Thanks for the first Phenom result for my list I just think you could have scored almost 2800 on SSE2...any applications running on the side ?

Yeah, I already made it run a bit 'slower', but even with under 10 seconds or the Core 2's running it at may be 5 seconds it delivers still pretty stable results...I could also do some zooming thing, but for the moment I focused more on the speed of the algoritm, so that all these new CPU's get something to do ;-) I'll try to keep up your ideas in the next releases...