Overview Features Coding ApolloOS Performance Forum Downloads Products Order Contact

Welcome to the Apollo Forum

This forum is for people interested in the APOLLO CPU.
Please read the forum usage manual.
Please visit our Apollo-Discord Server for support.



All TopicsNewsPerformanceGamesDemosApolloVampireAROSWorkbenchATARIReleases
Performance and Benchmark Results!

68K=Rules / PPC =Shitpage  1 2 3 4 5 6 7 

Gunnar von Boehn
(Apollo Team Member)
Posts 6207
17 Sep 2017 22:10


Markus Horbach wrote:

>APOLLO does beat all PowerPC clock by clock.
 
The benchmark results are a comparision with downclocked PowerPCs to get a clock2clock result ?

No of course not.
The benchmark lets each system run at their maximum full speed.

But we have to mind one thing here.
BUSTEST is the de facto standard memory benchmark.
Many AMIGA accelerators were advertised using this benchmark.

So we run the standard AMIGA tool to measure memory operations here.
This tool is an 68k program - as the vast majority of AMIGA programs are of course 68k programs.

So for 68K CPU this benchmark compares memory performance of the system.
For PPC system the benchmark includes as overhead the 68K emulation.

This is a drawback - but its also a realistic case as most AMIGA programs are 68k software.

So the question "how fast does this AMIGA program" run on this system - is a valid question also to ask a PPC system.


Samuel Devulder

Posts 248
17 Sep 2017 22:25


Ok, so the bench is not quite fair with full PPC machine. It mainly tests the memory speed of the 68k emulation. This is not the real memory speed of the system. This explain the huge difference with memory speed measured on mac-os.


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
17 Sep 2017 22:37


Samuel Devulder wrote:

    Ok, so the bench is not quite fair with full PPC machine. It mainly tests the memory speed of the 68k emulation. This is not the real memory speed of the system. This explain the huge difference with memory speed measured on mac-os.

 
First of all, you have to mind AMIGA OS is NOT MAC OS.
MAC OS was very well tuned to their HW.
You can not expect this from e.g. OS4.
   
In addition to this, mind that on AMIGA most programs are 68k code.
I thought this was obvious to all.
Like most AMIGA software - this  benchmark is 68K software.
The majority of real AMIGA software was of course written BEFORE PPC came out.
 
And also this standard AMIGA test is of course 68K code.
Same as Sysinfo and all the other typical AMIGA programs.
   
This test does answer the question:
Does my 2000 MHz PPC run this AMIGA 68K program faster than a real 68K running at 80 MHz.
 
And clearly no PPC is able to do this.

As we all recall, many years ago  when MAC switched to PPC, then
MAC did had the very same problem.
A lot MAC software was 68k.
And the new PPC did had to emulate the 68k to run the software.

Also the MAC had the same performance issues.
The first very low clocked PPC chip ran many MAC applications in very poor performance.

On the MAC side this changed quickly and later MACs ran 68k also very fast.
We see here a big difference to PPC AMIGA.


Samuel Devulder

Posts 248
17 Sep 2017 23:31


ok. no comment. (It's been deleted)
 


M Rickan

Posts 177
18 Sep 2017 02:37


Gunnar von Boehn wrote:

  Of course beating a ASIC which is 20 times higher clocked is something to expect is dreaming.
  Nevertheless APOLLO does even beat PPC systems being x20 higher in some benchmarks.

Since we're fantasizing:

How do you think the 080 in ASIC form would stack up against the PPC if clocked at the same rate?



Markus Horbach

Posts 35
18 Sep 2017 04:06


OK, 68k emulation on a PowerPC explains the huge differences. Now  I am curious about benchmarks of WinUAE on my Core i7-4770@3.4GHz. Or will this be unfair in some way ? Or can I expect a new world record ? A P96 benchmark might also be on my task list. Has someone already some WinUAE results ?


Markus Horbach

Posts 35
18 Sep 2017 04:12


>How do you think the 080 in ASIC form would stack up against the PPC if clocked at the same rate?
That mostly depends on the used external memory, not on the clock rate of the CPU core. If you clock the apollo core at lets say 1 GHz, the result of the memory bandwidth would be the same if you still use SDRAM with the same width and clock speed like the vampire V2.


Steve Ferrell

Posts 424
18 Sep 2017 04:31


Markus Horbach wrote:

  >APOLLO does beat all PowerPC clock by clock.
   
  The benchmark results are a comparision with downclocked PowerPCs to get a clock2clock result ?
 

 
Did you not read the benchmarks?  It states the clock speeds of the PPC CPU's and none of them were down clocked or under clocked as you've implied.  In case you missed them, here they are again:
 
 

      MOS IBook  PPC G4 1400 MHz  185.4 MB/sec
      MOS MacMini PPC G4 1500 MHz  184.3 MB/sec
      OS4 SAM PPC 460    1100 MHz  109.7 MB/sec
      OS4 SAM PPC 440    667 MHz    91.1 MB/sec
      OS4 Peg2 PPC G4    1000 MHz    80.6 MB/sec
 

 
 
 


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
18 Sep 2017 06:20


Markus Horbach wrote:

That mostly depends on the used external memory, not on the clock rate of the CPU core.

 
Actually, the performance of this test does depend more on the capabilities of the CPU than on the used memory.
 
Most of the 68K Systems use very similar memory.
Nevertheless they get totally different result.
 
The ACA Accelerator use the same memory as the Vampire.
But the Vampire scores 50 times better results than the ACA in the table. The reason is this is NOT the memory chip but the quality of the memory controller of the CPU.
 
 
Mind that all the PowerPC Systems in the list use faster memory modules than the VAMPIRE.
But the memory controller of most POWERPC are very bad.

PowerPC AMIGA can beat the VAMPIRE in this test.
Most of the PowerPC AMIGA memory controller can not match APOLLO- even with pure PPC code - simply because their memory controllers are  not of the same class.

 
The results printed in the table are a VAMPIRE-2, as all users have them. The VAMPIRE-4 uses faster memory than the VAMPIRE-2 and reaches even much higher scores.
 


Samuel Devulder

Posts 248
18 Sep 2017 07:59


Steve Ferrell wrote:
It states the clock speeds of the PPC CPU's and none of them were down clocked or under clocked as you've implied.  In case you missed them, here they are again:
   
   

        MOS IBook  PPC G4 1400 MHz  185.4 MB/sec
        MOS MacMini PPC G4 1500 MHz  184.3 MB/sec
        OS4 SAM PPC 460    1100 MHz  109.7 MB/sec
        OS4 SAM PPC 440    667 MHz    91.1 MB/sec
        OS4 Peg2 PPC G4    1000 MHz    80.6 MB/sec
   


Sorry, but PPC speed is totally meaningless since the benchmark was done using emulated 68k. The emulated 68k speed (say around 100mhz) would have been a much better indicator IMHO.

The big trouble with this benchmark is that it compares apples with oranges, plain 68k with emulated ones. These are two different matters.

Vampire is a nice accelerator, but better compare it with plain 680x0 accelerators. That would be much more meaningful.


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
18 Sep 2017 08:32


Samuel Devulder wrote:

  Sorry, but PPC speed is totally meaningless since the benchmark was done using emulated 68k.
 

   
Lets look at facts:
AMIGA software is mostly 68K, in fact 90% of AMINET software is 68K.
Therefore the question: how fast can AMIGA programs run, is a sensible question.
   
   
Another fact: PPC NG Systems suffer from low memory performance.
 
PEGASOS, AMIGA-ONE G3/G4, these PPC NG system reach, even with 100% PPC software, only halve of the performance of the VAMPIRE.
 
Recall the OS4 PPC STREAM results.
You will recall that the VAMP-2 scored 2 times better than the AMIGA-ONE.
And the AMIGA ONE was using native PPC code!

How can you explain that a 800 MHz PPC using native PPC code - scores less than 50% of the 80MHz 68K CPU ?


Steve Ferrell

Posts 424
18 Sep 2017 08:39


Sorry, but PPC speed is totally meaningless since the benchmark was done using emulated 68k. The emulated 68k speed (say around 100mhz) would have been a much better indicator IMHO.
The big trouble with this benchmark is that it compares apples with oranges, plain 68k with emulated ones. These are two different matters.

Vampire is a nice accelerator, but better compare it with plain 680x0 accelerators. That would be much more meaningful.

The results are hardly meaningless.  They're quite useful for those of us who are contemplating the purchase of a Vampire versus a PPC system.  I think you are too far down in the weeds so to speak....you can't see the forest because of the trees.


Samuel Devulder

Posts 248
18 Sep 2017 09:05


Steve Ferrell wrote:
The results are hardly meaningless.

  Well comparing emulation with real hardware is at least questionable.
 
  They're quite useful for those of us who are contemplating the purchase of a Vampire versus a PPC system.

  Right, but the available solutions are not restricted to (now outdated) G4-PPC machines. One should also consider 68k emulation on G5/Raspberry-PIs/PC as well.
   
Don't get me wrong, I don't like amiga-ppc for several reasons, but it's just a matter of being fair and comparing apple with apples, oranges with oranges. In my mind the title would be more accurate if it was "Softs on 68k = Rules / Softs on 68k-emulation-on-ppc = shit", because generally speaking, the performance of (obsolete)PPCs are honorable in many aspects (stream reports around 530MB/s for Apple_PwrMac_G4-500: EXTERNAL LINK )
   
But overall I think bustest is a bad test because memory bandwidth is only a part of computing power. One should also consider data-handling, hence instruction-set speed. Then I'm also pretty sure that the vampire gets on top of the list as well. Riva's results are good examples, more real-use, less marketting-numbers/hype.
   
Gunnar wrote:
How can you explain that a 800 MHz PPC using native PPC code - scores less than 50% of the 80MHz 68K CPU ?
Good point. Much more meaningful than comparing with emulated 68K IMHO. Let's compare HW with HW to be fair since this is all due to bad memory controllers of Amiga-PPCs.


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
18 Sep 2017 09:14


I agree with you that BUSTEST mixes two things.
Memory and being AMIGA software.
But with having a huge number of real 68K AMIGA software this is a common usecase.

But I agree with you is you want to understand the real problem of todays PPC AMIGAs then we need to look at them in detail.

Samuel Devulder wrote:

  "68k = Rules / 68k-emulation-on-ppc = shit", because generally speaking, the performance of (obsolete)PPCs are honorable in many aspects (stream reports around 530MB/s for Apple_PwrMac_G4-500: EXTERNAL LINK )
 

 
  But we look at AMIGA OS.
 
  Run stream on your favorite AMIGA PPC
  For example on AMIGA-ONE 800 MHz.
  EXTERNAL LINK 
  AMIGA-ONE using 800 MHz PPC = 153 MB/sec
 
The VAMP-2 can reach 2-3 times this value.
How can we explain why is the 800 MHz AMIGA ONE - several times slower than an 80 MHz 68K ?
 
And this AMIGA ONE is not a single example.

You can take the EFIKA

You can take the SAM

You can even take the latest X1000

and find benchmarks were all of them loose against an 80 MHz 68K CPU.


Samuel Devulder

Posts 248
18 Sep 2017 09:26


Gunnar von Boehn wrote:
 
  Run stream on your favorite AMIGA PPC
  For example on AMIGA-ONE 800 MHz.
  EXTERNAL LINK   
  AMIGA-ONE using 800 MHz PPC = 153 MB/sec
 
  The VAMP-2 can reach 2-3 times this value.
 

As stated above, this is a very good point in favor of the vampire. Very impressive! It should be emphasised over bustest because stream test is designed to compare different machines.
 
Don't get me wrong: I don't particularly like amiga-ppc designs as well, but let's be fair and compare HW with HW with proprer tools: 68080-code vs PPC-code. Stream benchmark is much more meaningful than bustest with that respect. Let's use bustest on 68k-based systems to rank them relatively, and use stream for cross-cpu comparison. IMHO.
 


Thellier Alain

Posts 141
18 Sep 2017 09:26


Hi
Vampire & Apollo are awesome 68k hardware ! Greeting to all the Apollo Team for bringing us the fastest 68k Amiga ever :-)

>AMIGA software is mostly 68K, in fact 90% of AMINET software is 68K.
Ok this is a respectable opinion: owning/using a PPC-Amiga  as a 68k accelerator for legacy RTG apps

Buut if you run (say) Lightwave 68k on an NG Amiga it will not use only the 68k emulation but also call the OS .libraries that are PPC and that use the NG hardware so the average speed will be better that if all was done 100% inside the 68k emulation (with the slow mem access that you describe well)

Alain Thellier




Steve Ferrell

Posts 424
18 Sep 2017 09:35


Samuel Devulder wrote:

Steve Ferrell wrote:
The results are hardly meaningless.

  Well comparing emulation with real hardware is at least questionable.
 
 
  They're quite useful for those of us who are contemplating the purchase of a Vampire versus a PPC system.

  Right, but the available solutions are not restricted to (now outdated) G4-PPC machines. One should also consider 68k emulation on G5/Raspberry-PIs/PC as well.
   
  Don't get me wrong, I don't like amiga-ppc for several reasons, but it's just a matter of being fair and comparing apple with apples, oranges with oranges. In my mind the title would be more accurate if it was "68k = Rules / 68k-emulation-on-ppc = shit", because generally speaking, the performance of (obsolete)PPCs are honorable in many aspects (stream reports around 530MB/s for Apple_PwrMac_G4-500: EXTERNAL LINK )
   
  But overall I think bustest is a bad test because memory bandwidth is only a part of computing power. One should also consider data-handling, hence instruction-set speed. Then I'm also pretty sure that the vampire gets on top of the list as well. Riva's results are good examples, more real-use, less marketting-numbers/hype.
 

Why do you keep referring to the Vampire as "emulation"?  The Vampire is no more an emulator than an EEPROM is an emulator.

And no one here is talking about comparisons or benchmarks of the Vampire versus UAE or other software emulators.  The benchmarks were for PPC systems running OS4 versus the Vampire running OS3. And I don't know of any G5's that run OS4 natively so why would I even care if the Vampire outperforms it?  If I wanted emulation, I'd just run WinUAE. And what would be the point of benchmarking a Vampire versus UAE?  That would actually be pointless.

And the benchmarks aren't marketing hype.  As I stated earlier, for those who are considering the purchase of an OS4 system versus an OS3 Vampire or other 68K acclerator, these benchmarks are invaluable.

If I'm in the market for hardware, why would I pay $2000 USD for a PPC Amiga running OS4 that is outlclassed by an FPGA costing a fraction of that price?

Remember, we're talking about hardware here, not benchmarking software emulators.  Since it seems so important to you, run some UAE/G5/Pi/PC benchmarks and post the results.  The results won't mean much to users here, because most forum members are interested in just 3 things:
1.  Upgrading classic hardware with a Vampire
2.  Upgrading classic hardware with a non-Vampire accelerator
3.  Considering the purchase of a Vampire vs a PPC Amiga

They're all very well aware of UAE and other emulation solutions.  Comparisons outside the scope above are pointless as no one here is looking for a Pi/G5/PC/UAE solution.




ExiE CZEX

Posts 48
18 Sep 2017 10:04


Steve Ferrell wrote:

And the benchmarks aren't marketing hype.  As I stated earlier, for those who are considering the purchase of an OS4 system versus an OS3 Vampire or other 68K acclerator, these benchmarks are invaluable.

Of course it is part of the marketing. And these synthetic benchmarks are hardy invaluable. Don't get me wrong, it is really nice that Vamp has so fast mem access, but it says nothing about real power of the machine.

Do you really believe people make their decisions on one test and not thinking about real power, available software, future possibilities etc.?

Steve Ferrell wrote:

They're all very well aware of UAE and other emulation solutions.  Comparisons outside the scope above are pointless as no one here is looking for a Pi/G5/PC/UAE solution.

So why compare Vamp with MOS running on Mac hardware?



Mr Niding

Posts 459
18 Sep 2017 10:13


ExiE CZEX wrote:

Steve Ferrell wrote:

  And the benchmarks aren't marketing hype.  As I stated earlier, for those who are considering the purchase of an OS4 system versus an OS3 Vampire or other 68K acclerator, these benchmarks are invaluable.
 

 
  Of course it is part of the marketing. And these synthetic benchmarks are hardy invaluable. Don't get me wrong, it is really nice that Vamp has so fast mem access, but it says nothing about real power of the machine.
 
  Do you really believe people make their decisions on one test and not thinking about real power, available software, future possibilities etc.?
 
 
Steve Ferrell wrote:

  They're all very well aware of UAE and other emulation solutions.  Comparisons outside the scope above are pointless as no one here is looking for a Pi/G5/PC/UAE solution.
 

  So why compare Vamp with MOS running on Mac hardware?
 

My take on it;

The more data refrence points you get for xyz hardware, the better a user can judge the merits of the options.
Ofcourse, it require an understanding of what the data actually translates to, and if there are software that will utilize the potential.


Samuel Devulder

Posts 248
18 Sep 2017 10:37


Steve Ferrell wrote:

  Why do you keep referring to the Vampire as "emulation"?
 

I don't.

posts 139page  1 2 3 4 5 6 7