Hi.

Out of curiosity, I created a tiny "benchmark" program that's calculating an approximation of the Pi number using the Monte Carlo method. The result is not particularly accurate (due to only 1000000 points in the code and the fact that the RndF() function in Hollywood does not generate number 1). However, it does what it's supposed to do - it overloads the computer with nonsense calculations and gives the result in seconds at the end.

To the point - somehow it's noticeably faster on the Mini. The G5 calculates the whole thing in about 24 seconds, while the G4 takes about 14 seconds. This is a huge difference, given the relative performance difference of these hardware.

I have a tiny request to all of you - could someone equipped with a computer with a G4 and/or G5 repeat these calculations? I have a sneaking suspicion that the binary generated in Hollywood is simply optimized for the G4 (presumably Andreas only has a Mac Mini), but I'd still like to make sure before I'll bother him about it.

EDIT: Oops, I might have uploaded bad binary - that's counting 100 million points. I'll do it again. Sorry.

[ Edited by koszer 08.03.2025 - 18:46 ]

Not sure what I should see here.

Both on the PowerBook and G5 I get a black "Hollywood" Window that eats all CPU without end (well over 1 minute) and simply stops when hitting the close gadget.

Quote:

You're right. I'm sorry about that. I've experimented earlier with more points to count - and I've uploaded not this binary.

Here's the proper one (that should complete calculations in less than half minute).

O.k. did you move the numbers the other way now?

G5 (AGP) 2.3GHz: 3.438s
PB5.9 1.67GHz: 6.329

What kind of G5 are you running?
Are you sure it runs at full clock? Maybe test something else like unpacking a big archive.

This has been a bug with at least the iMac a few years back (which I noticed when unpacking the beta ISO vs an 1.8GHz G4).

Quote:

Now it's calculating only 1 million points. On my G5 QUAD @ 2,5 GHz I get about 23 seconds while I get about 14 seconds on Mac Mini @ 1,33 GHz.

No other application running in the background, upgraded to MorphOS 3.19 not so long ago. How odd...
I'll give it a try from a live CD to rule out software fault.

EDIT: When I run the binary from "live USB" fresh MorphOS 3.19 I still get 23.158 seconds time.

[ Edited by koszer 08.03.2025 - 19:24 ]

Retried twice on the PB with todays's 3.20beta out of RAM: in shell I got 5.9s and 5.8s.

EFIKA on public 3.18 will do it in 26.7s kinda close to your G5 numbers.

The PB being faster than the Mini might be down to faster memory (in addition to +340MHz) but then your G5 should beat mine as DDR2 4200 vs DDR(1) 3200.

So again better check other benchmarks.

[ Edited by Kronos 08.03.2025 - 19:36 ]

The Altivec Unit of the G4 is faster than that of the G5. I don't know, if Holywood is using Altivec.

Quote:

Well, I just did dnetc benchmark on both machines. Here are the results:


Code:

The G5 is clearly faster (not by a huge margin in some cases, but still). What's going on? I'm baffled.

The iMac gets 23s which is inline with your Quad.

It kinda is an PCIe based PowerMac in a different package while the older unsupported 1st and 2nd generations would be releated to AGP G5s.

Quote:

Meanwhile zukow tried this binary on his PCIe G5@2,3 GHz and he got a 3.786s result.
Weirder and weirder...

PoweMac G5 2,7 Ghz : 24.582

Weird.

Quote:

Weird indeed.
What do we know so far?
Code:

This just doesn't add up. Why PPC970fx used in one machine would give similar result to 970MP used in another one, but so different from 970fx in yet another one?

I can confirm the results on Power Mac G5 PCIe 2,5 GHz (PPC970MP): 23,36s. I can test on another G5, this time AGP 2,3 GHz but first I must dust it off and update it to latest MOS.

Quote:

Thank you very much. At least I know my computer isn't broken (or I'm not the only one having a broken computer, heh).

I also got one result from A1139 (17 inch PowerBook G4 @ 1,67 GHz) and it's 13,67s. It's wild, as we've already got one 1,67 PowerBook G4 (but A1138 model as I believe) and Kronos reports it achieved 6,33s.

For what it's worth, 20.039s here on a Powerbook G4 15" 1,67ghz (A1138)...

Quote:

That's almost on par with slowest G5 results. No G4 has ever reported over 20 seconds.
What puzzles me is the fact that despite being quite random, these results seem strangely consistent on a specific machine - i.e. my G5 Quad gives me around 23 seconds no matter what tricks I use. I'll try disabling the CPU caches today.

I made some tests on Mac mini 1.5GHz and PB 1.67GHz (5,9 - A1139), and noticed quite some difference between OS versions on the very same machine, which may explain some results.

Mini with 3.19: 13.56s
Mini with 3.20: 5.23s
PB with 3.19: 13.72s
PB with 3.20: 5.88s

Funny that Mini is slightly faster than PB, but quite similar anyway. I'll try to test with some G5 machine(s) later in the evening or so.

BTW. can you make the test without using any Rnd functions, using other kind of calculations? Or share the source code for experiments?

Quote:

Interesting. Some bugfix maybe?

Quote:

As a matter of fact I did try some other calculations (pure adding in one instance) and Mini still won with G5 (not by such margin as in our disputed case, but nevertheless).
I even wrote a simple program drawing 200000 red boxes all over the screen. Still the G4 won with G5 (11 seconds vs 13 seconds). The only upside was - the executable generated for Windows did much worse, scoring well over 1 minute on my machine.

As for the code - nothing to be proud of, really. Here you go:


counter = 0

max_counter = 1000000

inside = 0

StartTimer(1)

For counter = 0 To max_counter

x = RndF()
y = RndF()

If (x^2+y^2) < 1 Then inside = inside+1

Next

epi = 4 * inside / max_counter

countingTime = GetTimer(1)

countingTime = countingTime / 1000

Print("przyblizona wartosc Pi to " .. epi .. ", obliczenia zajely " .. countingTime .. " sekund.")

WaitLeftMouse

End

Do the users ever get beta access? I assume not since i have never seen any betas of the OS available to us mere users.

[ Edited by ChrisC 10.03.2025 - 11:30 ]

Quote:

No, only system developers and beta-testers.

I've tried to determine if cpu cache has something to do with the results.
I typed CPU NOCACHE in shell and ran the tests again.
The results were identical as reported previously.

The iMac was still on a late 3.19 beta and after updating to 3.20 beta I get 4.38s.

I guess I have to get around to putting 3.19 onto the EFIKA to see if that "competitive" results stays.....

Quote:

You got 4,38s on the same machine that earlier gave you 23s by updating the operating system?
Now that's what I call an update!

So ...

I have dusted off my PMac G5 AGP 7,2 with 2x2,3 GHz PPC970fx, updated it from MOS 3.18 to 3.19 and ran the test: 24.66s. This is basically in-line with the PCIe ones since PPC970MP at 2,5 GHz is a bit faster.

Quote:

I made some further tests now.

First your Hollywood code can be optimized to much faster, and when doing that, results are more in line with the CPU performance.

I made a quick copypaste with your code, and it produces three lines now:
1) Your original code
2) Disabled Hollywood from doing some housekeeping in the heavy calculation loop (DisableLineHook)
3) Changed frequently used variables in the loop to Local instead of Global, which is always a good idea

I also ran the tests on PowerMac G5 2.7GHz. Here are the results:

Mini 3.19
przyblizona wartosc Pi to 3.141932, obliczenia zajely 13.537 sekund.
przyblizona wartosc Pi to 3.139372, obliczenia zajely 2.316 sekund.
przyblizona wartosc Pi to 3.144432, obliczenia zajely 1.425 sekund.

Mini 3.20
przyblizona wartosc Pi to 3.14024, obliczenia zajely 5.221 sekund.
przyblizona wartosc Pi to 3.1414, obliczenia zajely 2.296 sekund.
przyblizona wartosc Pi to 3.140884, obliczenia zajely 1.415 sekund.

PB 3.19
przyblizona wartosc Pi to 3.143672, obliczenia zajely 13.321 sekund.
przyblizona wartosc Pi to 3.141008, obliczenia zajely 2.163 sekund.
przyblizona wartosc Pi to 3.141592, obliczenia zajely 1.349 sekund.

PB 3.20
przyblizona wartosc Pi to 3.14244, obliczenia zajely 5.732 sekund.
przyblizona wartosc Pi to 3.142288, obliczenia zajely 2.194 sekund.
przyblizona wartosc Pi to 3.142924, obliczenia zajely 1.288 sekund.

G5 3.19
przyblizona wartosc Pi to 3.141284, obliczenia zajely 24.216 sekund.
przyblizona wartosc Pi to 3.141228, obliczenia zajely 1.283 sekund.
przyblizona wartosc Pi to 3.140952, obliczenia zajely 0.726 sekund.

G5 3.20
przyblizona wartosc Pi to 3.141824, obliczenia zajely 2.869 sekund.
przyblizona wartosc Pi to 3.141004, obliczenia zajely 1.258 sekund.
przyblizona wartosc Pi to 3.14116, obliczenia zajely 0.707 sekund.


So, it's basically only that unoptimized code on older MorphOS that is extra slow on G5 for some reason...

Here's my executable (prints to shell): ObliczPI3
And the code: ObliczPI3.hws.txt

> You got 4,38s on the same machine that earlier gave you 23s by
> updating the operating system? Now that's what I call an update!

That's not unheard of. With MorphOS 3.12 something similar happened with regard to Hollywood, as you probably remember:

https://morph.zone/modules/newbb_plus/viewtopic.php?forum=9&topic_id=12651