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!

Vampire2 80MHz Outclasses the AMIGAONE 800MHzpage  1 2 3 4 

Gunnar von Boehn
(Apollo Team Member)
Posts 6197
02 Sep 2017 22:02


Vampire2 80MHz outclasses the AMIGAONE 800MHz in the notorious memory benchmark STREAM.
 
  Let compare both system using the popular memory benchmark Stream
  EXTERNAL LINK 
  AmigaONE (512MB RAM, PowerPC 750FX @ 800Mhz, AmigaOS 4 Pre3)
  = Copy:  153.4 MB/sec
 
 
  Amiga600 Vampire2 78MHz AmigaOS 3.1
  = Copy:  319.9 MB/sec


Peter Heginbotham

Posts 214
02 Sep 2017 22:17


nice, any chance of posting the full output of both systems


Vojin Vidanovic

Posts 770
02 Sep 2017 22:32


Gunnar von Boehn wrote:

  AmigaONE (512MB RAM, PowerPC 750FX @ 800Mhz, AmigaOS 4 Pre3)
  = Copy:  153.4 MB/sec
  Amiga600 Vampire2 78MHz AmigaOS 3.1
  = Copy:  319.9 MB/sec

 
  Nice, fast mem is great. Surely it should with better mem controller and faster memory (DDR I suppose vs DDR2 on v2 Vamp and soon DDR3 on v4 Vamps). I remember
 
  Next thing: what will be ~clockrate of 080 in v4.
  That should be a good "class" improvement :-)
 
  Performance page
  CLICK HERE  needs also updates.
 
  Mind that ApolloWiki does too.
 


Gunnar von Boehn
(Apollo Team Member)
Posts 6197
02 Sep 2017 22:39


Vojin Vidanovic wrote:

    Nice, fast mem is great. Surely it should with better mem controller and faster memory (DDR I suppose vs DDR2 on v2 Vamp and soon DDR3 on v4 Vamps).

   
No, in theory the memory of the AMIGAONE is fastre than the memory of the Vampire2.
But only in theory.
   
The main reason is that POWERPC CPU and the AMIGAONE mainboard are unable to properly use the memory chips.
   
The reason for the better memory result is the quality of the CPU.
 
The Vampire is in real world much faster memory wise (Over twice as fast as the 800 MHz PPC) because the Apollo 68080 CPU is more capable than the PowerPC CPU.
 


Gunnar von Boehn
(Apollo Team Member)
Posts 6197
02 Sep 2017 22:43


Peter Heginbotham wrote:

nice, any chance of posting the full output of both systems

Yes on the V4

As you know STREAM does use not only stress the MEMORY but of course also the FPU.

On GOLD 2.7 only parts of the FPU are enabled in hardware.

On the V4 when all the FPU is enabled then the VAMPIRE will beat and clearly outclass the POWERPC AMIGA ONE also in all the other FPU stream tests.

Lets look forward to those results.


Thierry Atheist

Posts 644
02 Sep 2017 23:14


And the Vampire's AMMX will beat the AmigaOne's Altivec too, I assume?


Gunnar von Boehn
(Apollo Team Member)
Posts 6197
02 Sep 2017 23:21


Thierry Atheist wrote:

And the Vampire's AMMX will beat the AmigaOne's Altivec too, I assume?

Yes but it does depend on the algorithm.
AMMX2 is very strong but also very user and coder friendly.
AMMX2 is much nicer and easier to code than Altivec.

And actually most PPC AMIGA do not even have ALTIVEC. ;-)




Fernando Pereira

Posts 68
03 Sep 2017 11:50


Gunnar von Boehn wrote:

  Vampire2 80MHz outclasses the AMIGAONE 800MHz in the notorious memory benchmark STREAM.
 
  Let compare both system using the popular memory benchmark Stream
  EXTERNAL LINK   
  AmigaONE (512MB RAM, PowerPC 750FX @ 800Mhz, AmigaOS 4 Pre3)
  = Copy:  153.4 MB/sec
 
 
  Amiga600 Vampire2 78MHz AmigaOS 3.1
  = Copy:  319.9 MB/sec
 

 
  Nice! That's a considerable difference!
 


Gunnar von Boehn
(Apollo Team Member)
Posts 6197
03 Sep 2017 21:15


Update:

VAMPIRE-2 score increased to stream 319 MB/sec
So now over 2 times the score of the AMIGA ONE POWER-PC 800 MHz


Pedro Cotter
(Apollo Team Member)
Posts 308
03 Sep 2017 23:36


Eat my dust!!! :b
 
 


Rob M

Posts 60
04 Sep 2017 00:13


Vojin Vidanovic wrote:

  Nice, fast mem is great. Surely it should with better mem controller and faster memory (DDR I suppose vs DDR2 on v2 Vamp and soon DDR3 on v4 Vamps). I remember 

Articia S based hardware only supports 100 or 133Mhz SDR RAM. 


Gunnar von Boehn
(Apollo Team Member)
Posts 6197
04 Sep 2017 17:53


Rob M wrote:

 
Vojin Vidanovic wrote:

    Nice, fast mem is great. Surely it should with better mem controller and faster memory (DDR I suppose vs DDR2 on v2 Vamp and soon DDR3 on v4 Vamps). I remember 
 

 
  Articia S based hardware only supports 100 or 133Mhz SDR RAM. 
 

 
Thanks for the info.
The memory module would be 64bit?
 
So the expected peak speed in theory would be 133x8 = 1000 MB/sec
 
Comparing this to only reaching 150 MB/sec in Stream is shameful.
 
Calling such hardware "AMIGA" is an insult to the AMIGA name.


Rob M

Posts 60
05 Sep 2017 21:57


Gunnar von Boehn wrote:

Rob M wrote:

 
Vojin Vidanovic wrote:

      Nice, fast mem is great. Surely it should with better mem controller and faster memory (DDR I suppose vs DDR2 on v2 Vamp and soon DDR3 on v4 Vamps). I remember 
   

   
  Articia S based hardware only supports 100 or 133Mhz SDR RAM. 
 

  Thanks for the info.
  The memory module would be 64bit?
 
  So the expected peak speed in theory would be 133x8 = 1000 MB/sec
 
  Comparing this to only reaching 150 MB/sec in Stream is shameful.
 
  Calling such hardware "AMIGA" is an insult to the AMIGA name.

Yes PC66, PC100 and PC133 are all 64-bit wide.  According to JEDEC the specification supports a max of 1066MB/s, not sure what can realistically be expected but it's fair to assume the A1 is way off the mark.  Interestingly the CT60 is claimed to offer read and write speeds just of 100MB/s when the CPU and bus are clocked at 100Mhz.




Steve Ferrell

Posts 424
05 Sep 2017 22:13


Gunnar von Boehn wrote:

   
  Comparing this to only reaching 150 MB/sec in Stream is shameful.
   
  Calling such hardware "AMIGA" is an insult to the AMIGA name.
 

 
I feel the same way about current NG Amigas.  The benchmarks I've seen for those systems puts them on par with PC's that were on the market in 2003-2005. But is the poor performance of the Amiga One due to the hardware design or the OS or both?


Gunnar von Boehn
(Apollo Team Member)
Posts 6197
06 Sep 2017 08:50


Steve Ferrell wrote:

  But is the poor performance of the Amiga One due to the hardware design or the OS or both?
 

 
These Memory Speed tests are very low level.
This means they are small number of instructions and the OS is not directly used here.
 
The OS could influence these test with either "bad setup" of the memory controller, or bad usage/setup of the MMU.
If this is the case I can not answer.
 
 
Two things are clear.
 
There are two types of CPUs today.
a) very old CPUs or CPUs designed for low cost, low end markets
These CPUs have no means of detecting memory streaming, these CPUs are not designed or able to prefetch memory by themselves.
These CPUs are always slow with memory.
These groups includes:
  - ancient 68K chips of the 80th,
  - PowerPC 60x, PowePC 400-4xx, PowerPC 750, PowerPC 7400 ...
    So most chips use in PPC AMIGA
  - Old x86
 

The other group are
b) very modern, high end CPUs
These group includes new INTEL CORE chips, high end IBM POWER chips that only big companies can effort, and the APOLLO 68080
These cores detect memory streaming automatically and prefetch memory. These modern/high end chips play in regards of memory performance a different league than the old/low end chips.
 
 


Rob M

Posts 60
07 Sep 2017 20:33


Steve Ferrell wrote:

Gunnar von Boehn wrote:

   
    Comparing this to only reaching 150 MB/sec in Stream is shameful.
   
    Calling such hardware "AMIGA" is an insult to the AMIGA name.
 

 
  I feel the same way about current NG Amigas.  The benchmarks I've seen for those systems puts them on par with PC's that were on the market in 2003-2005. But is the poor performance of the Amiga One due to the hardware design or the OS or both?

I don't think the OS is the problem.  Obligement did a comparison between OS4.1 and MorphOS 2.3 on the Pegasos 2 and memory performance was pretty close, OS4 actually had the edge copying from RAM to RAM.  Also this test runs 27 faster on my X1000 than the A1XE G3 result.


Vojin Vidanovic

Posts 770
07 Sep 2017 20:49


Rob M wrote:

    Also this test runs 27 faster on my X1000 than the A1XE G3 result. 

 
  One of the good sides of PA Semi weird design is memory performance.
  2x2GB DDR2 (I believe it has interleave)

2013 UBUNTU 12.04 LTS A1-x1000 Phoronix results
  http://openbenchmarking.org/result/1310106-AR-AMIGAONE027#overview_table

  I dont know how does that compare to G5, and does it fall to old or new school.
 
  However, it does not save the day.
  And there is not much one can do to improve it (maybe low CL memory).
 
  As added value, one must change the RAM value in CFE on upgrades.
  EXTERNAL LINK   
  I am glad Vampire is a design that can grow in many ways.
 
 
Rob M wrote:

  Obligement did a comparison between OS4.1 2009 edition and MorphOS 2.3 on the Pegasos 2 and memory performance was pretty close, OS4 actually had the edge copying from RAM to RAM. 
 

 
  Overall, MorphOS did quite won, and with a lot of upadates since v2.3 tested, seems best thing x5000 users can hope is - MorphOS.
 
 
  EXTERNAL LINK  (click on US/British or your suitable Google Translate button.
Best viewed by French speakers)
 
  Conclusion
 
 
  With 228.5 points for MorphOS against 164 for AmigaOS 4.1, the comparison runs clearly to the advantage of MorphOS. If we take the averages, this represents 9.93 / 10 for the butterfly system and 7.13 / 10 for that with the bouncing ball. Virtually all domains are favorable to MorphOS in terms of performance (boot, USB, desktop, 3D, 68k emulation, WarpOS emulation and Warp3D, etc.).
 
  Contrary to what I had written at first, AmigaOS 4.1 on Pegasos II is not in beta. However, this is the first version of this system for Pegasos II. Will AmigaOS 4.x improve on this hardware? Maybe or maybe not. In the meantime, it remains less optimized than MorphOS. Finally, it may be noted that the performance of some tests is more due to the way the software was compiled than to the system itself.
 


Daytona 675x

Posts 32
11 Sep 2017 09:18


Very nice results for the Vampire, great work!
But...

I'm a Vampire fan. I got one myself and I love it.
But I am no fan of "beautified", imcomplete or simply wrong benchmarks.

Let's check out this one here and the claim you made.
I have no 800 MHz AONE, but I have a 667 MHz sam440ep.

This 667 MHz sam440ep "outclasses" the Vampire2 in this test, to use your wording :)

sam440ep, 667 MHz:
Function      Rate (MB/s)  Avg time    Min time    Max time
Copy:        497.3274      0.0649      0.0643      0.0686
Scale:        326.4240      0.0987      0.0980      0.1031
Add:          323.2801      0.1494      0.1485      0.1536
Triad:        334.3107      0.1444      0.1436      0.1489

Let's ignore the FPU-calc parts of this benchmark (although, the Vampires current absolute lack of performance in those areas is one of the reasons why e.g. a T57 version is simply not possible on the V2 in its current state, no matter what some people still continue to say / hope).

Note that I rebuilt the PPC stream-exe with a much newer GCC (using -O3 as the readme says, of course). Obviously the antique build in the archive you linked to didn't do a good job. Here's what I get with that old build (only copy-speed differs significantly, the others are practically identical):

sam440ep, old stream-exe from 2005
Copy:        329.7538

Still enough to beat the V2 by a bit ;) But a huge difference!

If you want to do benchmark masturbation then I suggest you make sure you actually measure somewhat correctly. Obviously the old stream-exe doesn't unleash even near the copy-speed those PPCs are truely capable of if using the tools (e.g. dcbt) those CPUs provide.

Here's the sam460ex result btw.:
Copy (new exe): 782.0543
Copy (old exe): 427.5389

You should probably remeasure the AONE results you posted before making toooooo strong claims about the speed of certain PPC systems here...
Here's a link to the rebuild:
EXTERNAL LINK 
Speaking about fairness of your benchmarks:

just like Vojin I noticed that the performance page here needs to be reworked:
CLICK HERE 
It shows a wrong value for the sam440ep at least. I made the sortbench test on the sam440ep some months ago when you asked so in the a1k forum and I got 299.01, not 250.11. Which in turn means that the sam440ep scores higher than the Vampire, unless the value of the Vampire grew in the meantime.
EXTERNAL LINK 
Also: sortbench spits out more results than just the 32k test. Those others are just as likely to be seen in real-world programs as the 32k test (actually those smaller ones are probably much more likely).
Why is it that we aren't informed that a sam440ep "outclasses" the V2 in the 1k to 8k tests by about factor 3?
(considering the numbers from half a year ago).

I mean, if you're doing benchmarks, then please don't just mention the partial results that are in favor of the Vampire. Such misleading information is one of the reasons why some people get unrealistic expectations (like with T57 :P ).

The Vampire is such a great thing, you really shouldn't taint its reputation by "beautifying" benchmark-results or by trying to "bash" PPC (at least that's the impression one might get here very easily).

There's really no need to, we all know that the Vampire is a wonderful piece of work. Let's just stay fair, reasonable and down on earth :)


Gregthe Canuck

Posts 274
11 Sep 2017 09:47


Hi Daytona -

Nice to see you posting in this forum.

Yes benchmarks are in many cases as much a comparison of compilers as anything else... so hopefully everyone takes them for what they are.

Looking forward to your future contributions here. There is a lot of ongoing work on the Gold 2.7 and 3 cores so today's benchmark could be obsolete tomorrow.

I think you will find the Vampire very interesting as a possible target for a Tower57 port... or at least I'm hoping... ;)

Cheers!




Gunnar von Boehn
(Apollo Team Member)
Posts 6197
11 Sep 2017 09:56


Daytona 675x wrote:

The Vampire is such a great thing, you really shouldn't taint its reputation by "beautifying" benchmark-results or by trying to "bash" PPC (at least that's the impression one might get here very easily).

 
The POWER-PC result were "official" OS4 results.
We did not invent them but the OS4 developer included his result with the AMINET upload.
 
 
Daytona 675x wrote:

I suggest you make sure you actually measure somewhat correctly. Obviously the old stream-exe doesn't unleash even near the copy-speed those PPCs are truely capable of if using the tools (e.g. dcbt) those CPUs provide

 
STREAM is an official benchmark.
Your "handwriting the code" is not allowed to STREAM benchmark guidelines.
Your "handwriting" is regarded as cheating, and does not give result which are officially accepted.
 
 
We ran the benchmark on the VAMPIRE as compiled by the "old" GCC 2.9
And we gave the scores as it.
And the Vampire scores were as a matter of fact 2 times higher than the scores of the OS4 developer for the AMIGA-ONE

If we handtune the code on 68k - then we could also reach much higher scores.
 
In fact, on the V2 with handwritten stream, we reached over 600 MB/sec.
And on the V4 over 1000 MB/sec stream score.
 
So beating your handtuned SAM score is also no problem for the Vampire. But we did not post these "handtuned" scores.


posts 70page  1 2 3 4