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!

Atari Powapage  1 2 

Gunnar von Boehn
(Apollo Team Member)
Posts 6207
08 Apr 2018 14:37


Vampire GOLD 2.9 running as ATARI
 

 
VDI GFX driver is not optimized for SAGA yet.
We can expect significant speedups, if people help to tune the driver.
 
IDE Driver is also not tuned yet.
 


Vojin Vidanovic
(Needs Verification)
Posts 1916/ 1
08 Apr 2018 15:37


Very nice results, esp. regarding high end Ataris such as 040 upgraded Milan or Falcon - in CPU, FPU, 3D it blows the "Atari(STs) mind".

Hope tuned up MINT installation can be oficially distributed, since its free (and most of Atari software seems to be abamdonware by now)
 



Gunnar von Boehn
(Apollo Team Member)
Posts 6207
08 Apr 2018 15:45


Vojin Vidanovic wrote:

  Very nice results, esp. regarding high end Ataris such as 040 upgraded Milan or Falcon - in CPU, FPU, 3D it blows the "Atari(STs) mind".
 

 
An overclocked 68060 at 100MHz plus SuperVidelGFX card is certainly a nice and fast system.

Nevertheless the Vampire has no problem to win every benchmark against the overclocked 68060 + SuperVidel.

The only reason that the Vamp does not win some benchmarks today is
that the 68060+Videl uses handtuned ASM drivers - while VAMP today uses generic untuned drivers in C.
 
As soon as there are ASM tuned drivers for VAMP/SAGA the benchmarks result will be very clear.


Vojin Vidanovic
(Needs Verification)
Posts 1916/ 1
08 Apr 2018 16:51


Gunnar von Boehn wrote:

  As soon as there are ASM tuned drivers for VAMP/SAGA the benchmarks result will be very clear.
 

 
  Surely, hardware implementation is superior enough. I am glad there is a plan to optimize drivers since FreeMINT distros seems to be quite unoptimized. I have found a 040 optimized kernel e.g.
  freemint-1.15.12-kernel-040.tar.gz
  Could Vamp benefit from it? I hope in future 080 can be written, and with optimization of XAES/TeraDesk seems its all finetuning it would need (which is real nice compared to AROS, AmiKit ...)

Sidenote:
Its interesting to see EmuTOS is now in the core :-)
That is its highest known implementation (or lowest from h/w side)
EXTERNAL LINK 


Olivier Landemarre

Posts 147
08 Apr 2018 17:56


Gunnar von Boehn wrote:

Vojin Vidanovic wrote:

  Very nice results, esp. regarding high end Ataris such as 040 upgraded Milan or Falcon - in CPU, FPU, 3D it blows the "Atari(STs) mind".
 

 
  An overclocked 68060 at 100MHz plus SuperVidelGFX card is certainly a nice and fast system.
 
  Nevertheless the Vampire has no problem to win every benchmark against the overclocked 68060 + SuperVidel.
 
  The only reason that the Vamp does not win some benchmarks today is
  that the 68060+Videl uses handtuned ASM drivers - while VAMP today uses generic untuned drivers in C.
 
  As soon as there are ASM tuned drivers for VAMP/SAGA the benchmarks result will be very clear.

Better and better results, yes looks like faster than 68060 CT60 but for a reason I not understand the reference take here is far not the best result I have (I have no results with supervidel I'm going to ask to a friend) but with Radeon CT60 give better result of around +50% (without take into account GFX as Radeon driver is hardware accelerated). If you could provide the CT63 result (your default.abh this could be nice for me I do some collection of benchs!) and of course your last Vamp results! But normal results on CT60 100Mhz can be find here : EXTERNAL LINK  replace your default.abh by this file and you will have good better idea how better is Vamp (but no doubt Vamp is faster!)
Don't forgot CPU and FPU are stupid tests as superscalar can't be used so it is plus minus link to processor frequency.
You have done very nice job, this is quite incredible.
Do you plane add some speed up hardware for GFX such memory fill?

Thanks



Gunnar von Boehn
(Apollo Team Member)
Posts 6207
08 Apr 2018 18:09


Olivier Landemarre wrote:

Do you plane add some speed up hardware for GFX such memory fill?

Right now the limitation is purely the drivers and the software.

For example KRONOS does some benchmarks on PLANAR mode in ST-RAM.
Our SAGA can display ATARI PLANAR GFX modes and put out on DIGITAL-VIDEO,
but so far EMUTOS does not use this but use old AMIGA chipset for this. SAGA is about 50-100 times faster than the old AMIGA chipset.

So relative small change to EMUTOS /ATARI to use SAGA for PLANAR modes will result in huge speedup.

The VDI driver on SAGA is a generic C code.
A tuned driver doing the work loop in tuned ASM can for sure double the scores in VDI benchmarks.

APOLLO is very fast in memory operations.
The CPU can do in Memory fill 700-800 MB/sec.
The limit is not the CPU but the driver.




Ingo Uhlemann

Posts 35
08 Apr 2018 18:35


Vojin Vidanovic wrote:

Gunnar von Boehn wrote:

    As soon as there are ASM tuned drivers for VAMP/SAGA the benchmarks result will be very clear.
 

 
  Surely, hardware implementation is superior enough. I am glad there is a plan to optimize drivers since FreeMINT distros seems to be quite unoptimized. I have found a 040 optimized kernel e.g.
  freemint-1.15.12-kernel-040.tar.gz
  Could Vamp benefit from it? I hope in future 080 can be written, and with optimization of XAES/TeraDesk seems its all finetuning it would need (which is real nice compared to AROS, AmiKit ...)
 
  Sidenote:
  Its interesting to see EmuTOS is now in the core :-)
  That is its highest known implementation (or lowest from h/w side)
  EXTERNAL LINK 

Surely Freemint has special compiled kernel for 020, 030, 040,060 cpu. Freemint Version is 1.19 and the Freemint what is used on the vampire is newest compiled with 040 optimizings.
 


Vojin Vidanovic
(Needs Verification)
Posts 1916/ 1
08 Apr 2018 19:34


Ingo Uhlemann wrote:

  Surely Freemint has special compiled kernel for 020, 030, 040,060 cpu. Freemint Version is 1.19 and the Freemint what is used on the vampire is newest compiled with 040 optimizings. 
 

 
  I am glad its already 040 optimized. Some speed could be gained with additional 080 compile, when possible. I am glad EmuTOS officially supports Vampire and MINT too ... Nice union.

Gunnar von Boehn wrote:

  So relative small change to EMUTOS /ATARI to use SAGA for PLANAR modes will result in huge speedup.
 
  APOLLO is very fast in memory operations.
  The CPU can do in Memory fill 700-800 MB/sec.
  The limit is not the CPU but the driver.

No doubt, with these optimizations, Vamp implementation will stand out :-)


OneSTone O2o

Posts 159
10 Apr 2018 18:43


You should get in touch with Thorsten Otto, who is one of the guys in charge of maintaining gcc for Atari plattform. So 68080 optimisation could be implemented.


Sebastian Blanco

Posts 148
10 Apr 2018 21:13


This is so cool :D


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
11 Apr 2018 18:05


Vampire
 
  197.9 ATARI KRONOS BogoMIPS
 
 
  Is this any good compared to other 68K or Coldfire?


OneSTone O2o

Posts 159
11 Apr 2018 18:50


That bogomips not bad, but not far from my Falcon 030 with ct60e at 16 Mhz bus clock and 96 Mhz CPU... 188,94...
   
EXTERNAL LINK     
EXTERNAL LINK     
EXTERNAL LINK 
(how to integrate pictures here, html&[img] does not work)
     
I think Firebee with Coldfire still has higher bogomips, by price of partial incompatibility.
     


Markus (mfro)

Posts 99
11 Apr 2018 19:19


I think everybody agreed already that BogoMips values are basically useless?
 
  They more or less just measure how long it takes a certain CPU to do nothing ;).
 
  Just for completeness: FireBee BogoMips score is about 265 (probably a slight Kronos rounding/measuring inaccuracy as the FireBee is clocked at 264 MHz).


Vojin Vidanovic
(Needs Verification)
Posts 1916/ 1
11 Apr 2018 19:20


oneSTone o2o wrote:

That bogomips not bad, but not far from my Falcon 030 with ct60e at 16 Mhz bus clock and 96 Mhz CPU... 188,94...

Current average X11 Vamp @80Mhz should rate as 060 100Mhz. Overclocked X12 cores a bit more.
Surely, with V4s clock could go to X13 or X15 I hope, as well as more cache, faster RAM and full FPU should contribute a bit. From there, next gen Vampire with bigger FPGA would be needed for more, as in any other case.



Gunnar von Boehn
(Apollo Team Member)
Posts 6207
11 Apr 2018 19:27


Markus (mfro) wrote:

  Just for completeness: FireBee BogoMips score is about 265 (probably a slight Kronos rounding/measuring inaccuracy as the FireBee is clocked at 264 MHz).
 

Thanks!

I still need to learn how to configure Kronos.
What results does Firebee score in CPU and FPU chart compared to Falcon?


Markus (mfro)

Posts 99
11 Apr 2018 19:37


Gunnar von Boehn wrote:
  What results does Firebee score in CPU and FPU chart compared to Falcon?
 

 
  Not really comparable as Kronos is only available as m68k executable.
 
  Thus the FireBee has to mill some of the benchmarks through its integrated emulator while others can be executed natively.
 
  Here are some numbers with an early version of the emulator: EXTERNAL LINK


Sebastian Blanco

Posts 148
11 Apr 2018 19:39


Very impressive.


Olivier Landemarre

Posts 147
11 Apr 2018 20:16


Kronos is 68K but test are fully coldfire, as tests are inside dynamic library.

Olivier

Markus (mfro) wrote:

Gunnar von Boehn wrote:
  What results does Firebee score in CPU and FPU chart compared to Falcon?
   

   
    Not really comparable as Kronos is only available as m68k executable.
   
    Thus the FireBee has to mill some of the benchmarks through its integrated emulator while others can be executed natively.
   
  Here are some numbers with an early version of the emulator: EXTERNAL LINK 




Gunnar von Boehn
(Apollo Team Member)
Posts 6207
11 Apr 2018 20:55


Olivier Landemarre wrote:

  Kronos is 68K but test are fully coldfire, as tests are inside dynamic library.
 

 
  Ah yes I found the Firebee results.
  And the test code is indeed Coldfire "native/clean"
 
 
 
              Coldfire    68080      Milan-68040
  CPU_BitShift  =  265      198        12
  CPU_RawPower  =  268      200        24
 
  CPU_MEMOREAD  =    83      534        34
  CPU_MEMOWRITE =    47      663        21 
  -----------------------------------------
 

 
  pretty interesting
 
 
 


Olivier Landemarre

Posts 147
12 Apr 2018 18:41


Gunnar von Boehn wrote:

Olivier Landemarre wrote:

  Kronos is 68K but test are fully coldfire, as tests are inside dynamic library.
 

 
  Ah yes I found the Firebee results.
  And the test code is indeed Coldfire "native/clean"
 
 
 
                Coldfire    68080      Milan-68040
  CPU_BitShift  =  265      198        12
  CPU_RawPower  =  268      200        24
 
  CPU_MEMOREAD  =    83      534        34
  CPU_MEMOWRITE =    47      663        21 
  -----------------------------------------
 

 
  pretty interesting
 
 
 

Results for firebee are quite old, I'm going to check if they do some progress since, nobody was able to understand why memory access was so low while processor speed is quite fast if run in cache, it was the same when I was using coldfire evaluation board from freescale, perhaps some wrong cache management.
On several 68060 systems results could change a lot with configuration, I know a lot of that have used Kronos on CT60 just to fix the configuration up they was able to acheive reference results!



posts 24page  1 2