Amiga Emulation Benchmarks

[minator]

I recently got interested in the Amiga again ...because I'd bought a new Mac.
I bought a new machine because I wanted to run local Ai models. To run the bigger models takes an enormous amount of video memory and if you want to go above 48GB, Macs with Apple silicon work out to be the cheap option. Of course, Apple being Apple don't just let you spec out any machine with a large memory, for the top memory you have to get the top processor and that's $$$.

What's any of this to do with the Amiga?

This thing is ridiculously fast. Apart from running Ai, which maxes out the GPU, I don't run anything that goes anywhere near the power of the machine. So it got me thinking about how good would it be at emulation, and that got me interested in the Amiga again.

I fired up an emulator yesterday and ran some benchmarks:

Emulation System:
CPU: M3 Max 4GHz (16 Core CPU / 40 Core GPU)
RAM: 128GB

Emulator: Amiberry preview 6.3.3 macOS-apple-silicon
JIT: No (not available for ARM64).

Emulation settings:
68040 - fastest.
AGA chipset - fastest.
2MB Chip, 4MB Fast RAM.

Amiga OS:
Kickstart 3.1
Workbench 3.1 - The Workbench I used was a backup I had made from the Walker.

SysInfo 4.4
471.12 (x A600)
249 227 Dhrystones
260 MIPS
60 MFlops


AIBB 6.5
Test code type: 68020, 40 Math
Results are KoopRate (A600 is 1.00)


EmuTest 266
WritePixel 340
Sieve 310
Drystone 293
Sort 242

EllipseTest 8.82
Matrix 470
IMath 1864
MemTest 563 (677 MB / Sec)
TGTest 5.66

LineTest 2.25
Savage 3017
FMath 1463
FMatrix 478
BeachBall 3224

InstTest 335
Flops 4326
TranTest 2869
FTrace 2601
CplxTest 385



Conclusion:
This is pretty impressive considering this is a pre-release non-JIT emulator and Amiberry appears to use all of 1 core.

Memory speed was a bit disappointing considering the external RAM bandwidth on this machine is 400GB / Second and this emulation was likely sitting entirely in the cache.


PiStorm [ Show youtube player ] gets mostly faster results on a considerably slower CPU so I'd really like to see this when a JIT emulator is implemented in Amiberry.
I'd also like to see the emulator split across 2 cores. It should be possible to run the CPU and Fast RAM emulation separately from the chipset so I suspect that'd provide a performance boost.

If this is how fast it can be emulated, imagine how fast it'd be running native?
The A600GS is running some Arm native parts so maybe we'll find out.

[Mr-Z]

PiStorm also has the advantage it only has to do 68K CPU emulation.
So not having to emulate the Amiga chipset saves cycles

[stx2199]

for emulator, games and most programs what really matter is single core speed
and the single core speed of the M3 is impressive acording to passmark benchmark
in on the level of the core i9 14900ks and 14900kf which are the faster x86 cpus today
https://www.cpubenchmark.net/singleThread.html



but that's according to passmark
in real life the things are different, and I really don't believe in passmark results when mix x86 and arm cpus, because is obvious that the x86 cpus are much faster in single core than the arm ones


OK ,your numbers posted indicate that such Amiga emulator is faster than winuae without JIT enabled
but winuae on any old x86 cpu with JIT enabled is between 5x and 10x faster than what you got

anyways interesting results thx for posting them

[minator]

Originally Posted by stx2199

for emulator, games and most programs what really matter is single core speed
and the single core speed of the M3 is impressive acording to passmark benchmark
in on the level of the core i9 14900ks and 14900kf which are the faster x86 cpus today
https://www.cpubenchmark.net/singleThread.html


but that's according to passmark
in real life the things are different, and I really don't believe in passmark results when mix x86 and arm cpus,

Those benchmarks are weird, they're showing M3 as faster than M4, that makes no sense as the M4 is a more advanced core and 400MHz faster.

because is obvious that the x86 cpus are much faster in single core than the arm ones

Why do you think that?
Granted most Arms are designed for phones so wont be as fast as desktop cores, but the Apple ones are designed to be fast. They also have faster decode than x86 and most importantly get the fastest silicon process before anyone else.

OK ,your numbers posted indicate that such Amiga emulator is faster than winuae without JIT enabled
but winuae on any old x86 cpu with JIT enabled is between 5x and 10x faster than what you got

So that's good. I've seen a package? called Amikit for the RPi5 that does 2500 MIPS in SysInfo so that that makes sense. That'll be using an Arm32 JIT.

It'll be interesting to see once Amiberry gets an Arm64 JIT how my scores turn out. Should be quite a bit better than the RPi5.

anyways interesting results thx for posting them

AIBB gave some very curious results depending on the code type / mode. CP-Math (co-processor?) was much faster for floating point. Not sure what was going on there.

[stx2199]

Originally Posted by minator

Why do you think that?
Granted most Arms are designed for phones so wont be as fast as desktop cores, but the Apple ones are designed to be fast. They also have faster decode than x86 and most importantly get the fastest silicon process before anyone else.

plz note the apple m3 and m4 are fast but no even close to a core i7 or i9 or even the i3 or any other modern desktop x86 cpu
the m3 and m4 are just 30w cpus while the core i9 i7 waste +200w when are pushed
yep the arm cpus are more efficient but they still have a long way to go to compare to x86 power

The ppl of passmark and geekbench put numbers on par in single core with the x86 but it is just a marketing strategy to avoid ppl realizing that the arms cpus are actually much slower

[Seiya]

i think you should try to make benchmark with real software.
Many years ago i make a benchmark 68k between Amithlon (athlonXP 2600, Sempron 2400, Pentium 2), AmigaOne SE G3, Pegasos G3, Pegasos II G4

[minator]

Originally Posted by stx2199

plz note the apple m3 and m4 are fast but no even close to a core i7 or i9 or even the i3 or any other modern desktop x86 cpu
the m3 and m4 are just 30w cpus while the core i9 i7 waste +200w when are pushed

Power consumption is not a measure of performance.

yep the arm cpus are more efficient but they still have a long way to go to compare to x86 power

If that was the case Apple would still use x86.
It's also not my experience. When I first got an M1 Pro, it completely destroyed my i7, and it doesn't get hot in normal use.

The x86 machines are still faster on the desktop because they have more cores and have higher clock speeds, but the M3 max isn't that far behind.

The ppl of passmark and geekbench put numbers on par in single core with the x86

Passmark didn't make sense to me at first but I think it's not quite measuring the same thing as Geekbench. Geekbench is more of a system test, so things like cache size and memory bandwidth matter.
Passmark appears to be more of a CPU only test, the M4 is faster than an M3 (it's a more advanced core and on an improved process). However it's also in an iPad so it more likely to throttle if it's pushed, that makes sense.

but it is just a marketing strategy to avoid ppl realizing that the arms cpus are actually much slower

No benchmarks are perfect but I've no reason to believe they are being manipulated.
Arm were relatively basic low end chips 15 years ago, but a lot has changed since then.

[PeterK]

Don't overvalue the results of simple benchmarks, the speed of real applications is something different.

The Jit can improve speed a lot. On my 5 year old low-cost Celeron N4100, 4 core CPU, with WinUAE 4.4 and Jit on, my results are much higher. Without Jit the benchmark results are poor, but for applications the differences are lower. Btw, you can enable "multithreaded RTG" in WinUAE for a faster graphics emulation.

[Seiya]

you can try this serie of benchmark (download files)

Some tests are very slow without jit. So you can do with JIT and without JIT. The only things to remember is to tell if you using jit and/or in what test you use jit.

CPU TEST:

ByteMARK 68k/PPC
AFlops only mflops(4)
Dnet68k (v2.8010-461-CPR-00062713) (I provide the 68k archive)
Amiga Fibonacci Benchmark (Afinch)


AUDIO TESTS:

BladeENC (0.85 for 68k Amiga
Lame 3.98

GAMES and ENTERTAINMENT:

QUAKE (software mode): use Quake68k (included in the archive) and than "timedemo demo2" from the console by pressing the "" key under the "Esc" key and then pressing it again immediately to see the demo in fullscreen; Or download the Glquake.
QUAKE II (software mode).


How to run the benchmarks:

BYTEMARK: try the versions for your cpu
AFLOPS: test that calculates the mflops with these options: aflops (using the right version for your cpu); we only take into account MFLOPS(4)
DNET 68K (in this case): launch the program from shell and write: "dnet -benchmark" and report the RC5 and OGR value.
AMIGA FIBONACCI TEST: open the shell and launch the right program for the cpu version
BLADEENC: use the wav file i included from Demoscene, indicated at the bottom of the page and from the shell do: "BladeEnc Osaka_Theory.wav"
LAME 3.98: always with the same song, this time using qval=3 so always from the shell: "lame -q3 Osaka_Theory.wav Osaka_Theory.mp3"
QUAKE 68k: simply do a "timedemo demo2" from the console being careful to immediately press the "" key under the "esc" key to restore fullscreen and at the end bring back the correct frame rate.
Test quake with AGA and RTG . Resolution to try: 320x256x8 (AGA), 640x480x8 (RTG).
In RTG mode run the game from shell with this command line: "quake68k -ip 127.0.0.1" and then choose 640x480 8bit
QUAKE 2 ("timedemo 2" then "map demo1.dm2"): possibly the same and without audio. Important for the purposes of the tests that the games are tested in software mode


audio test i used Demoscene music (Osaka_Theory.wav) free from copyright (i presume).

[mschulz]

Originally Posted by Seiya

you can try this serie of benchmark (download files)

Fantastic! Let me use this to measure Emu68 too (once I have it running again )