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!

FPU VAMPIRE2 Beat All Available AMIGA Acceleratorspage  1 2 3 

Vojin Vidanovic

Posts 770
04 Sep 2017 17:41


Samuel Devulder wrote:

  So I would'nt say that sysinfo results are fake per se. They just give an indication of performance for non highly hardware-optimized code, i.e. average programs.

They are comparable to other same version Sysinfo results and that is it. And that is also where Vamp is so far superior, even it doesnt use a single of its advanced feats.

I wonder why dont they use Minibench
EXTERNAL LINK


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
04 Sep 2017 17:48


Samuel Devulder wrote:

  So I would'nt say that sysinfo results are fake per se.
 

 
  SYSINFO is 100% fake.
 
  SYSINFO calls their results  "Dhrystones" - but does bot implement the dhrystone code at all.
 
  And based on the fake "Dhrystone" value the MIPS is calculated.
 
  This is like you running 2 times around your block taking the time and then saying you did a marathon run in time xx.
 


Samuel Devulder

Posts 248
04 Sep 2017 18:05


Right, it's only the label that is misleading (or fake if you prefer :) ). But Sysinfo still gives valid information on the expected speed one can get from a vampire with an average-joe 20year-old program I guess. And that is very interresting as these old progs are unlikey to be de-compiled and re-optimized any soon.


Szyk Cech

Posts 191
04 Sep 2017 18:13


Instead complain on SysInfo maybe you can propose some other more representative benchmark?!? With real MIPS and dhrystone. I suppose that such benchmark exist...


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
04 Sep 2017 18:19


Samuel Devulder wrote:

Right, it's only the label that is misleading (or fake if you prefer :) ). But Sysinfo still gives valid information on the expected speed one can get from a vampire with an average-joe 20year-old program I guess. And that is very interresting as these old progs are unlikey to be de-compiled and re-optimized any soon.

As you know the number of programs you can write is infinite.

So how can you compare a CPU speed?
Typically benchmarks therefore implement some meaningful routines like:
- sorting an array
- calculating prime numbers
- decoding a JPEG
 
This allows you to compare performance in doing realistic jobs.

Sysinfo doesjust some non functional nonsense code.
And print the result out under the "official" labels like Dhrystone and MIPS. This is fake and the Sysinfo scores do NOT reflect real Dhrystone or real MIPS values.

Of course in AMIGA land SYSINFO is known and most people are not aware of what SYSINFO does.
There are other, better benchmarks.
AIBB and MINIBENCH are a lot more meaningful then SYSINFO IMHO


Samuel Devulder

Posts 248
04 Sep 2017 18:34


Point taken. But yet SYSINFO is still a de facto standard for the amiga that is hard to get rid off. Even here on appolo-core forum, one can see more screenshots of SYSINFO than of AIBB (just look at the first message in this thread. Puzzling! ;-) )


Vojin Vidanovic

Posts 770
04 Sep 2017 18:53


Samuel Devulder wrote:

Point taken. But yet SYSINFO is still a de facto standard for the amiga that is hard to get rid off. Even here on appolo-core forum, one can see more screenshots of SYSINFO than of AIBB (just look at the first message in this thread. Puzzling! ;-) )

I wonder why dont they use Minibench
EXTERNAL LINK 
Point scoring is easer to comprehend and to meassure progress
between cores. Sysinfo just has nice info screen.


Vojin Vidanovic

Posts 770
05 Sep 2017 09:04


Chris Dennett wrote:

  Yes, on the v2 femu is patched into the kickstart and loaded before any FPU detection to trap FPU instructions not implemented in the core and show the Amiga as having an FPU without having to do any manipulations later on. Totally transparent to the user.

Exact things that should be kickstarted with Vampire.
Great work!


Thierry Atheist

Posts 644
05 Sep 2017 10:44


Gunnar von Boehn wrote:
The Sysinfo code wrongly not uses the 2nd Pipe on Super Scalar CPUs like 68060 and 68080.
There for a 68060@50 scores basically only halve the Dhrystone and MIPS numbers in SYSINFO - compared to the real official results from MOTOROLA.

Basically for 68060 and 68080 you can assume to x2 the score shown in SYSINFO to get a result close to a real Dhrystone score.


Hi Gunnar,

Correct me if I'm wrong, but that means that, if AOS1.x, 2.x and 3.x (if it were possible) were recompiled, with no enhancements except for directly related to the 68080 CPU would be, significantly faster, maybe 50 to 70%?!?!!!!!! :-D :-D :-D


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
05 Sep 2017 11:35


Thierry Atheist wrote:

  Correct me if I'm wrong, but that means that, if AOS1.x, 2.x and 3.x (if it were possible) were recompiled, with no enhancements except for directly related to the 68080 CPU would be, significantly faster, maybe 50 to 70%?!?!!!!!! :-D :-D :-D
 

AMIGA OS, and the majority of all old software is good and makes good usage of 2 pipes already.
 
Its just SYSINFO which is crappy and giving fake values.


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
05 Sep 2017 19:24


Update:
more FEMU performance tuning:




Steve Ferrell

Posts 424
05 Sep 2017 20:20


Nice work!


Peter Heginbotham

Posts 214
05 Sep 2017 20:29


nice :)

Is this getting to 040 performance levels ?


Samuel Devulder

Posts 248
05 Sep 2017 20:47


Peter Heginbotham wrote:
Is this getting to 040 performance levels ?

68040@40Mhz gives around 7.66MFlops
[img]http://eab.abime.net/attachment.php?attachmentid=27731&d=1296481575[/img]
at 25Mhz, it is 4.74MFlops. We're getting closer and closer each time.




Samuel Devulder

Posts 248
05 Sep 2017 20:47


Peter Heginbotham wrote:
Is this getting to 040 performance levels ?
 

  68040@40Mhz gives around 7.66MFlops ( EXTERNAL LINK ) and at 25Mhz, it is 4.74MFlops ( EXTERNAL LINK ).
 
 


Rod March

Posts 119
06 Sep 2017 00:04


Just keeps getting better :D


Gunnar von Boehn
(Apollo Team Member)
Posts 6207
06 Sep 2017 14:29


UPDATE:

FEMU tuning for VAMPIRE-2

   

   
We can see major speed improvements after tuning FEMU for 68080.
 
Mind, that this FEMU version can only run on 68080,
it uses extra registers only available on 68080
special instructions and 64bit operations only available on APOLLO 68080.



Peter Heginbotham

Posts 214
06 Sep 2017 14:41


so at this rate of improvment 7 Mflops this time tomorrow then :)


Michal Warzecha

Posts 209
06 Sep 2017 16:44


But it's score from SYSinfo, can anyone test it on some "real" benchamrk like LightWave?


Vojin Vidanovic

Posts 770
06 Sep 2017 16:51


Michal Warzecha wrote:

  But it's score from SYSinfo, can anyone test it on some "real" benchamrk like LightWave?
 

 
  ... which is comparable to SysInfo.
 
  I am more for MiniMon and yes, some real world performance test.
 
  As apetizer, I advise Impluses Imagine basic render scene with candle.  Used to render for hours on expanded A1200.
 
 
Peter Heginbotham wrote:

  so at this rate of improvment 7 Mflops this time tomorrow then :)
 

 
  Not really, and no, softFPU will not beat highly clocked 040 and 060. But at least will bring decent 8882 like FPU performance.
  But hardFPU might.


posts 54page  1 2 3