Poster | Thread |
Samwel
| |
Re: Benchmarks Posted on 17-Jan-2006 13:48:27
| | [ #41 ] |
|
|
|
Elite Member |
Joined: 7-Apr-2004 Posts: 3404
From: Sweden | | |
|
| @Chain-Q
The GX is faster than 7447!! That's a fact you can't force two CPU's down to 133MHz and say that one is faster in same circumstance. They should be tested with optimal conditions. With the right north bridge (200MHz bus, like TSi108) I'm quite sure it will out run 7447. Again excluding altivec.
But of course I agree that a G4 or any model that has altivec is prefered over the G3's. It's a shame that Troika didn't choose a G4 for their board. The latest specs and with a 7447/7447A @+1GHz would have been a really good OS4 motherboard IMHO.
_________________ /Harry
[SOLD] µA1-C - 750GX 800MHz - 512MB - Antec Aria case
Avatar by HNL_DK! |
|
Status: Offline |
|
|
itix
| |
Re: Benchmarks Posted on 17-Jan-2006 15:52:33
| | [ #42 ] |
|
|
|
Elite Member |
Joined: 22-Dec-2004 Posts: 3398
From: Freedom world | | |
|
| @Samwel
Uhm, that is like saying 68060 is faster than G4 at 50MHz (8KB L1 caches, without L2 cache and same bus)... But until 68060 runs at 5000MHz it is irrelevant...
Anyhow how much faster GX is compared to G4 at same clock speed and same cache size? Is it just faster on some operations or in everything excluding Altivec?
Last edited by itix on 17-Jan-2006 at 03:55 PM. Last edited by itix on 17-Jan-2006 at 03:53 PM.
_________________ Amiga Developer Amiga 500, Efika, Mac Mini and PowerBook |
|
Status: Offline |
|
|
wegster
| |
Re: Benchmarks Posted on 17-Jan-2006 16:39:20
| | [ #43 ] |
|
|
|
Elite Member |
Joined: 29-Nov-2004 Posts: 8554
From: RTP, NC USA | | |
|
| @Samwel
Quote:
The GX is faster than 7447!! That's a fact you can't force two CPU's down to 133MHz and say that one is faster in same circumstance. |
A fact should then be borned out by benchmarks. I'd expect the larger cache of the FX to compensate for a bit of clock speed, anywhere that cache hits can be heavily used, but unsure regarding simple computations where the cache isn't used often.
Quote:
They should be tested with optimal conditions. With the right north bridge (200MHz bus, like TSi108) I'm quite sure it will out run 7447. Again excluding altivec. |
Currently, a 200MHz bus isn't relevant, as we don't have systems with a 200MHz bus, only 133MHz. As this is Amiga and related content, it's much more interesting/useful/appropriate to 'test what we've got' IMO.
Quote:
But of course I agree that a G4 or any model that has altivec is prefered over the G3's. It's a shame that Troika didn't choose a G4 for their board. The latest specs and with a 7447/7447A @+1GHz would have been a really good OS4 motherboard IMHO. |
Yep, although it might have increased cost. Whatever the reason, hopefully the next version/model...
_________________ Are we not done with the same silly arguments and flames yet??! |
|
Status: Offline |
|
|
umisef
| |
Re: Benchmarks Posted on 19-Jan-2006 1:02:24
| | [ #44 ] |
|
|
|
Super Member |
Joined: 19-Jun-2005 Posts: 1714
From: Melbourne, Australia | | |
|
| @Seehund
Quote:
OK, dnetc would be a benchmark then. How good or useful a benchmark I wouldn't know. |
As a benchmark for emulation, it's about as useless as useless can be. Especially as a benchmark for recompiling (JIT, dynamic, call it what you want) emulation.
And even *if*, for some bizarre reason, one wanted to compare 68k emulations using 68k dnetc, for the love of deity, make sure you are at least running the same CODE! If one runs the 020 core, and one the 060 core, then the results are even more meaningless than usual...
Why is dnetc useless as an emulation benchmark? Well, speaking for RC5-64, the 040 core's main, inner loop
* Has 400+ instructions in a row, unbroken by any jump or conditional * Does not have a single instruction that operates on anything smaller than 32 bit * Does provably not examine any flags for the vast majority of those 400+ instructions, meaning 68k flag generation can be suppressed. * Has bugger all memory accesses, and those it has are nicely spaced to avoid any chance of turnaround latency, and are perfectly naturally aligned
Normal application code, on average, has between 3 and 6 instructions before encountering a (usually conditional, or calculated), is heavy with partial register accesses, looks at flags all the time (or at least can't be proven not to look at flags), so while *some* flag generation can be avoided, a lot can't. Oh, and memory accesses are all over the place, non-aligned almost as often as aligned.
A 68k JIT emulator's Nirvana is code having the first three listed properties of the RC5-64 core. Where JIT emulators slow down is when having to deal with jumps and conditionals. On PPC, dealing with less-than-32-bit operations is also a pain, involving multiple instructions to do even the simplest thing. And generating 68k flags (which are no less irregular than x86 flags) on a PPC (which has a very simple, very regular flag generation scheme) will slow things right down. Especially for the rotate instructions used in RC5, for which the 68k flags are even more convoluted than the x86 ones....
If you insist on running dnetc as an "emulation benchmark", then at least force each machine to use the 020 core; That way, the JIT will use something a little(!) more like real application code...
Of course, the interpretive emulation can't reap most of those benefits, because it looks at one instruction at a time. It *ALWAYS* has to create flags. It *always* treats the address of the next instruction as an unknown surprise. |
|
Status: Offline |
|
|
umisef
| |
Re: Benchmarks Posted on 19-Jan-2006 1:40:48
| | [ #45 ] |
|
|
|
Super Member |
Joined: 19-Jun-2005 Posts: 1714
From: Melbourne, Australia | | |
|
| @thread
As a suggestion --- 4.5 years ago, I published a list of benchmarks, then run on Amithlon, with detailed instruction on how to run them on the users' own machines. All of them are 68k, and all of them are freely available (and the descriptions mention where they can be obtained). And most of them are more sensible than running dnetc :)
I have just uploaded the textfile again, so if people are relly interested in comparing things (and especially if there is someone out there with Petunia running), it might be a good idea to try and reproduce as many of those as possible. Especially the bzip2 one is nice (replace the data file mentioned in the text with http://www.kernel.org/pub/linux/kernel/v1.0/linux-1.0.tar.bz2). |
|
Status: Offline |
|
|
wegster
| |
Re: Benchmarks Posted on 19-Jan-2006 1:58:12
| | [ #46 ] |
|
|
|
Elite Member |
Joined: 29-Nov-2004 Posts: 8554
From: RTP, NC USA | | |
|
| @umisef Thanks for the insight, as well as the text file...I _have_ to laugh at this quote from it, though
(RE: dnetc) Quote:
Yes, I *know* these are not benchmarks. I know these shouldn't even be considered for benchmarking. But of course, some people *do* use them for benchmarking (you know who you are!), so I will provide numbers. Unless you feel an irresistible urge to compare machines based on their RC5 or OGR scores, you should completely ignore this section! |
_________________ Are we not done with the same silly arguments and flames yet??! |
|
Status: Offline |
|
|
Samwel
| |
Re: Benchmarks Posted on 19-Jan-2006 2:37:28
| | [ #47 ] |
|
|
|
Elite Member |
Joined: 7-Apr-2004 Posts: 3404
From: Sweden | | |
|
| @itix
That wasn't really what I meant.
I meant that you shouldn't compare a 200MHz FSB CPU @ 800MHz with a 167MHz FSB CPU @ 1GHz connected to DIFFERENT northbridges (and ram) which runs @ 133MHz FSB. This seriously hurt the GX performance. Also "calculating" up the speed from 800MHz to 1GHz while running @133MHz also limits the gain GX would have made.
As I recall it most operations where faster (excluding altivec optimized routines of course). Due to the MUCH larger cache of 1MB and the higher FSB. But 7448 blows the GX away totally of course
* Note this is not applicable on the µA1-C as it has the Articia S which only run FSB @ 133MHz and not use the GX fully. Any PegasosII or A1-XE would be faster than this implementation.
_________________ /Harry
[SOLD] µA1-C - 750GX 800MHz - 512MB - Antec Aria case
Avatar by HNL_DK! |
|
Status: Offline |
|
|
tomazkid
| |
Re: Benchmarks Posted on 8-Feb-2006 22:19:40
| | [ #48 ] |
|
|
|
Team Member |
Joined: 31-Jul-2003 Posts: 11694
From: Kristianstad, Sweden | | |
|
| |
Status: Offline |
|
|
umisef
| |
Re: Benchmarks Posted on 9-Feb-2006 15:30:48
| | [ #49 ] |
|
|
|
Super Member |
Joined: 19-Jun-2005 Posts: 1714
From: Melbourne, Australia | | |
|
| @tomazkid
Quote:
[Feb 08 22:15:21 UTC] Automatic processor type detection found a Motorola 68020 processor. [Feb 08 22:15:21 UTC] RC5-72: using core #0 (MH 1-pipe 68020/030). [Feb 08 22:15:40 UTC] RC5-72: Benchmark for core #0 (MH 1-pipe 68020/030) 0.00:00:17.07 [811,721 keys/sec] [Feb 08 22:15:40 UTC] OGR-P2: using core #1 (GARSP 6.0 68020). [Feb 08 22:15:58 UTC] OGR-P2: Benchmark for core #1 (GARSP 6.0 68020) 0.00:00:16.19 [2,414,877 nodes/sec]
|
To be honest, I find this a rather disappointing result.
When a G4 emulating a 68k is not only slower than a 2001 Athlon emulating a 68k in absolute terms (812kkeys/s vs 2.2Mkeys/s), but also per clock cycle (just over 1 key per kilocycle, vs just over 1.5 keys per kilocycle), then that's a bit sad. Same for OGR (2.4Mnodes/s vs 3.47Mnodes/s, but at least here the per-MHz looks better --- 3 nodes per kilocycle, vs 2.6 nodes per kilocycle).
Of course, the 2001 figure were for much older clients, and running RC5-64 (which, however, as far as I can tell from statistics, runs at the same keyrate as rc5-72), so that, too, should be in the G4's favour (together with its 32 general purpose registers, and it being big-endian). Not to mention that the G4 doesn't have to emulate hardware, and the Athlon did.
Worse yet, on this particular RC5 "benchmark" (see my earlier posts :), Petunia on a G4/800 in task-based emulation seems to do a measly 10% better than my 333MHz iMac using my "3 month spare time hobby" PPC-JIT, and about 80% faster on OGR; In RC5, my JIT on a 600MHz G3 gets 60% *more* performance than Petunia on the G4/800, and almost exactly the same on OGR.
I would have expected more from an emulator that doesn't need to worry about hardware emulation, and runs on a nice RISC processor...
|
|
Status: Offline |
|
|
tomazkid
| |
Re: Benchmarks Posted on 9-Feb-2006 15:39:37
| | [ #50 ] |
|
|
|
Team Member |
Joined: 31-Jul-2003 Posts: 11694
From: Kristianstad, Sweden | | |
|
| @umisef
The rc5-72 actually went to about 1,100,100 keys/sec when I used the cores for 060 and 040 instead of the 020/030 , and OGR improved a little as well, but not that much.
_________________ Site admins are people too..pooff! |
|
Status: Offline |
|
|
Zylesea
| |
Re: Benchmarks Posted on 14-Feb-2006 10:04:47
| | [ #51 ] |
|
|
|
Elite Member |
Joined: 16-Mar-2004 Posts: 2263
From: Ostwestfalen, FRG | | |
|
| @Zylesea
Quote:
Zylesea wrote: @tomazkid
I did some benches comparing Peg1, BlizzardPPC, WinUAE and 68040. Site is in German, but numbers and configs are clear anyway.
http://via.i-networx.de/bench_en.html (English) http://via.i-networx.de/bench.html (German)
List is vey short, but gives a brief estimate between the 68k, MOS on the Blizzard and Peg and UAE.
edit: added english link |
Now that update4 with JIT is out, anyone with an (u)A1 willing to share numbers on benoit and Cinema4D? All my test parameters are listed at the link. C4D Version is 1.6.1998 (the copy that was sold withteh German mag "Amiga spezial"._________________ My programs: via.bckrs.de MorphOS user since V0.4 (2001) |
|
Status: Offline |
|
|
Mr_Capehill
| |
Re: Benchmarks Posted on 14-Feb-2006 11:38:35
| | [ #52 ] |
|
|
|
Super Member |
Joined: 15-Mar-2003 Posts: 1933
From: Yharnam | | |
|
| @umisef
I'm curios about what that IMac is running, is it "Amithlon PPC" or something like that? Haven't seen/heard before.
|
|
Status: Offline |
|
|
spotUP
| |
Re: Benchmarks Posted on 22-Feb-2006 16:29:16
| | [ #53 ] |
|
|
|
Elite Member |
Joined: 19-Aug-2003 Posts: 2896
From: Up Rough Demo Squad | | |
|
| @Chain-Q
hmmm.. my a1/xe/g3 renders that picture in 4-5 secs now. :) (petunia upd4)
edit: uhm, what resolution did you use btw?
Last edited by spotUP on 22-Feb-2006 at 04:35 PM.
_________________ AOS4 Betatester, Peg2, G4@1ghz, Radeon 9250 256mb, 1gb RAM.
http://www.asciiarena.com http://www.uprough.net |
|
Status: Offline |
|
|
tomazkid
| |
Re: Benchmarks Posted on 26-Dec-2006 3:29:30
| | [ #54 ] |
|
|
|
Team Member |
Joined: 31-Jul-2003 Posts: 11694
From: Kristianstad, Sweden | | |
|
| @tomazkid
with Update 3 Quote:
dnetc v2.9010-495-CTR-05051421 for AmigaOS (OS 4.0pre, 68K).
[Jan 13 03:03:39 UTC] Automatic processor type detection found a Motorola 68020 processor. [Jan 13 03:03:39 UTC] RC5-72: using core #0 (MH 1-pipe 68020/030). [Jan 13 03:03:59 UTC] RC5-72: Benchmark for core #0 (MH 1-pipe 68020/030) 0.00:00:16.69 [41,970 keys/sec] [Jan 13 03:03:59 UTC] OGR-P2: using core #1 (GARSP 6.0 68020). [Jan 13 03:04:19 UTC] OGR-P2: Benchmark for core #1 (GARSP 6.0 68020) 0.00:00:17.83 [186,412 nodes/sec]
This with 7450/51 at 800 mhz. |
With Update 4 Quote:
[Feb 08 22:15:21 UTC] Automatic processor type detection found a Motorola 68020 processor. [Feb 08 22:15:21 UTC] RC5-72: using core #0 (MH 1-pipe 68020/030). [Feb 08 22:15:40 UTC] RC5-72: Benchmark for core #0 (MH 1-pipe 68020/030) 0.00:00:17.07 [811,721 keys/sec] [Feb 08 22:15:40 UTC] OGR-P2: using core #1 (GARSP 6.0 68020). [Feb 08 22:15:58 UTC] OGR-P2: Benchmark for core #1 (GARSP 6.0 68020) 0.00:00:16.19 [2,414,877 nodes/sec] |
OS4 final
dnetc v2.9010-495-CTR-05051421 for AmigaOS (OS 4.0pre, 68K).
[Dec 26 03:17:09 UTC] Automatic processor type detection found a Motorola 68020 processor. [Dec 26 03:17:09 UTC] RC5-72: using core #0 (MH 1-pipe 68020/030). [Dec 26 03:17:27 UTC] RC5-72: Benchmark for core #0 (MH 1-pipe 68020/030) 0.00:00:16.29 [827,822 keys/sec] [Dec 26 03:17:27 UTC] RC5-72: using core #1 (MH 1-pipe 68000/040). [Dec 26 03:17:46 UTC] RC5-72: Benchmark for core #1 (MH 1-pipe 68000/040) 0.00:00:16.13 [1,221,837 keys/sec] [Dec 26 03:17:46 UTC] RC5-72: using core #2 (MH 2-pipe 68060). [Dec 26 03:18:05 UTC] RC5-72: Benchmark for core #2 (MH 2-pipe 68060) 0.00:00:17.07 [1,325,471 keys/sec] [Dec 26 03:18:05 UTC] OGR-P2: using core #0 (GARSP 6.0 68000). [Dec 26 03:18:23 UTC] OGR-P2: Benchmark for core #0 (GARSP 6.0 68000) 0.00:00:16.30 [2,080,608 nodes/sec] [Dec 26 03:18:23 UTC] OGR-P2: using core #1 (GARSP 6.0 68020). [Dec 26 03:18:42 UTC] OGR-P2: Benchmark for core #1 (GARSP 6.0 68020) 0.00:00:16.40 [2,503,624 nodes/sec] [Dec 26 03:18:42 UTC] OGR-P2: using core #2 (GARSP 6.0 68030). [Dec 26 03:19:00 UTC] OGR-P2: Benchmark for core #2 (GARSP 6.0 68030) 0.00:00:16.40 [2,503,689 nodes/sec] [Dec 26 03:19:00 UTC] OGR-P2: using core #3 (GARSP 6.0 68040). [Dec 26 03:19:20 UTC] OGR-P2: Benchmark for core #3 (GARSP 6.0 68040) 0.00:00:16.91 [2,489,382 nodes/sec] [Dec 26 03:19:20 UTC] OGR-P2: using core #4 (GARSP 6.0 68060). [Dec 26 03:19:39 UTC] OGR-P2: Benchmark for core #4 (GARSP 6.0 68060) 0.00:00:16.94 [2,503,869 nodes/sec]
Last edited by tomazkid on 26-Dec-2006 at 03:37 AM. Last edited by tomazkid on 26-Dec-2006 at 03:35 AM. Last edited by tomazkid on 26-Dec-2006 at 03:32 AM.
_________________ Site admins are people too..pooff! |
|
Status: Offline |
|
|
tomazkid
| |
Re: Benchmarks Posted on 22-Jan-2008 15:33:24
| | [ #55 ] |
|
|
|
Team Member |
Joined: 31-Jul-2003 Posts: 11694
From: Kristianstad, Sweden | | |
|
| BUMP
Now when OS4 Classic is out, perhaps you OS4 Classic users could add some benchmarks ?
_________________ Site admins are people too..pooff! |
|
Status: Offline |
|
|
tomazkid
| |
Re: Benchmarks Posted on 19-May-2008 14:32:47
| | [ #56 ] |
|
|
|
Team Member |
Joined: 31-Jul-2003 Posts: 11694
From: Kristianstad, Sweden | | |
|
| Bump!
_________________ Site admins are people too..pooff! |
|
Status: Offline |
|
|
DBAlex
| |
Re: Benchmarks Posted on 19-May-2008 15:50:06
| | [ #57 ] |
|
|
|
Cult Member |
Joined: 23-Jul-2006 Posts: 756
From: UK | | |
|
| @tomazkid
Won't they just be the same as Morphos being used on a PPC classic machine?
_________________ A1200, 68060/64MB/1.2GB/WiFi/AGAtoCRT/OS3.9 Pegasos I, G3 600Mhz/512/9200SE/80GB WinUAE, Ryzen 5 2400G/Vega11, 8GB DDR4, 256GB SSD,Win 10 Pro x64 Amiga Forever ! |
|
Status: Offline |
|
|
tomazkid
| |
Re: Benchmarks Posted on 19-May-2008 17:17:34
| | [ #58 ] |
|
|
|
Team Member |
Joined: 31-Jul-2003 Posts: 11694
From: Kristianstad, Sweden | | |
|
| @DBAlex
This is regarding 68k emulation mostly. _________________ Site admins are people too..pooff! |
|
Status: Offline |
|
|
NutsAboutAmiga
| |
Re: Benchmarks Posted on 19-May-2008 21:21:22
| | [ #59 ] |
|
|
|
Elite Member |
Joined: 9-Jun-2004 Posts: 12818
From: Norway | | |
|
| |
Status: Offline |
|
|