Amithlon Benchmarks
In the following, I have compiled the results for a range of easily
accessible benchmarks. I hope that this may give some idea of the relative
performance you can expect from an Amithlon system.
I have not commented any of the results; I want people to be able
to form their own, objective opinion. There are some "outliers" in there
(i.e. values that are far away from the "usual" and "expected" outcomes).
Be careful to check for the influence of such values on any averages you
might look at.
Also, I have occasionally taken the ratios of arithmetic averages;
That is not really a good thing to do, but if I don't, people will complain
and/or suggest a hidden motive. It is better than taking arithmetic averages
of ratios, at least....
I hope this will come in useful to someone,
Bernie (bmeyer@csse.monash.edu.au)
%%%%%%%%%%%%
% BYTEMARK %
%%%%%%%%%%%%
This is running the Bytemark port straight from Aminet (see
http://www.aminet.net/pub/aminet/util/moni/ByteMark68kPPC.lha
for your own copy). Running nbench.040fpu on Amithlon. The following
are the results as reported on a 1.3GHz Athlon. The comparison values
were
taken from Usenet post
<37893DA5.MD-0.198.s.giovenella@spamfree.pn.nettuno.it>
and were measured on an A3000/CSPPC with 66 MHz 060 and 233 MHz 604e.
The comparative MHz columns are not to be taken too seriously, because
performance will not actually increase linearly with CPU speed; Memory
bandwidth is also an important factor, and thus an 060 or 604e overclocked
to these frequencies would still not actually get anywhere near these
speeds.
BYTEmark (tm) Native Mode Benchmark ver. 2 (3/95)
Test ! Amithlon
1.3GHz ! 060-66 ! 604e-233 ! 060MHz ! 604MHz
---------------------------------!----------------------------------------
NUMERIC SORT: Index: 7.360227 ! 0.65
! 3.21 ! 747.3 ! 534.2
STRING SORT: Index: 6.692935 ! 0.47
! 1.28 ! 939.8 ! 1218.3
BITFIELD: Index: 6.756898
! 1.05 ! 4.58 !
424.7 ! 343.7
FP EMULATION: Index: 7.119398 ! 0.87
! 3.33 ! 540.0 ! 498.1
FOURIER: Index: 3.872222
! 0.18 ! 0.79 ! 1419.8
! 1142.0
ASSIGNMENT: Index: 12.250218 ! 0.94
! 4.77 ! 860.1 ! 598.3
IDEA: Index:
8.842450 ! 0.65 ! 4.99
! 897.8 ! 412.8
HUFFMAN: Index: 10.073270
! 0.93 ! 4.74 !
714.8 ! 495.1
NEURAL NET: Index: 3.154568 ! 0.35
! 2.60 ! 594.8 ! 282.6
LU DECOMPOSITION:dex: 3.849416 ! 0.54
! 2.15 ! 470.4 ! 417.1
===========OVERALL============= ! ======== ! ======== !
======= ! ======
INTEGER INDEX: 8.242951
! 0.77 ! 3.55 !
706.5 ! 541.0
FLOATING-POINT INDEX: 3.609326 ! 0.32
! 1.64 ! 744.4 ! 512.7
=============================== ! ======== ! ======== !
======= ! ======
%%%%%%%%%%%%%%%
% Speedometer %
%%%%%%%%%%%%%%%
This is Speedometer 4.02, running under Shapeshifter 3.10.
Screenmode is 1024x768 in 256 colours. The system this is run on is
a 1.3GHz
Athlon Thunderbird, with a GeForce2 MX-200 graphics card.
Comparison values for 060/66 were taken from
http://www.amigapro.com/Images/Speedometer.gif
Comparison values for iFusion on CSPPC 604e-200 were taken from
http://www.blittersoft.com/speed.jpg
Once again, don't pay too much attention to the comparative MHz figures,
because even if one could overclock the 060 or 604e to these levels,
one
would still be quite short of the actual performance observed under
Amithlon.
The "Disk" test was left out, because Speedometer 4.02 is faulty in
the way it reports relative disk speeds larger than 5[1]
Test ! Amithlon
1.3GHz ! 060-66 ! 604e-200 ! 060MHz ! 604MHz
---------------------------------!----------------------------------------
CPU
! 22.78 !
4.24 ! 9.71 ! 354.5
! 469.2
Math !
405.78 ! 44.47 ! 393.90
! 602.2 ! 206.0
...........................................................................
KWhet !
74.05 ! 12.90 ! 312.10
! 378.8 ! 47.4
Dhry !
36.65 ! 2.88 !
12.15 ! 839.8 ! 603.2
Towers !
38.23 ! 3.55 !
14.20 ! 710.7 ! 538.4
Quicksort !
36.12 ! 3.62 !
17.37 ! 658.5 ! 415.8
Bubble Sort !
37.04 ! 3.79 !
11.83 ! 645.0 ! 626.2
Queens !
39.35 ! 3.16 !
10.34 ! 821.8 ! 761.1
Puzzle !
39.06 ! 4.71 !
20.16 ! 547.3 ! 387.5
Permute !
33.96 ! 2.63 !
13.44 ! 852.2 ! 505.3
Int. Matrix !
43.82 ! 6.44 !
18.64 ! 449.0 ! 470.1
Sieve !
45.39 ! 4.09 !
14.25 ! 732.4 ! 637.0
Bench Average ! 42.37
! 4.78 ! 44.45 !
585.0 ! 190.6
...........................................................................
FPU FFT !
36.20 ! 3.29 !
2.09 ! 726.2 ! 3464.1
FPU KWhet !
30.69 ! 4.05 !
26.31 ! 500.1 ! 233.2
FPU Matrix !
31.64 ! 4.25 !
23.55 ! 491.3 ! 268.7
FPU Average !
32.85 ! 3.86 !
17.32 ! 561.6 ! 379.3
...........................................................................
8 Bit GFX !
4.73 ! 2.05 !
n/a !
%%%%%%%%%%%%%%%%%%%%%%%%%%
% Candy Factory Pro Demo %
%%%%%%%%%%%%%%%%%%%%%%%%%%
This is CandyFactoryPro, Demo version, straight from Aminet. See
http://www.aminet.net/pub/aminet/biz/titan/CandyProDemo.lha
for your own copy.
Once again, this is running on a 1.3GHz Athlon Thunderbird. Screen mode
is 1024x768 in 256 colours.
Comparison values for 604e-233 were taken from
http://www.amigapro.com/benchmarks.html
All timings in seconds!
Test ! Amithlon
1.3GHz ! 604e-233 ! 604MHz
---------------------------------!----------------------------------------
BlueGlow !
0.12s ! 0.16s ! 310
Choccy !
0.15s ! 0.17s ! 264
ChromeShadow ! 0.09s
! 0.08s ! 207
NiceBrown !
0.17s ! 0.18s ! 246
PurpleGoo !
0.50s ! 0.93s ! 433
ShinyMetal !
0.47s ! 1.23s ! 609
Tinfoil !
0.11s ! 0.09s ! 190
Transparent !
0.09s ! 0.08s ! 207
%%%%%%%%%%%%
% P96Speed %
%%%%%%%%%%%%
The gfx speedtest program available on Aminet at
http://www.aminet.net/pub/aminet/gfx/board/P96Speed.lha
060-70/A4000T/PIV comparison values for 640x480x8 were taken from the
included archive.
060-50/Mediator/Voodoo3 2000 comparison values for 640x480x8 and 800x600x16
were taken from message
on the Yahoo Mediator mailing list.
This is again running on a 1.3GHz Athlon Thunderbird, with a Geforce2
MX-200
(That's the cheapest and slowest card in the GeForce2 range, at a current
street price of about US$ 50-60)
Test (640x480x8) ! Amithlon ! 060-70/PIV
! 060-50/Mediator/V3-2000
-------------------------------------------------------------------------
RectFill !
20,071 ! 6,330
! 19,093
RectFillPat. ! 3,772
! 5,911 !
17,152
WritePixel ! 1,930,706
! 174,161 !
147,037
WriteChunkyP. ! 11,599
! 685
! 590
WritePixelArr. ! 11,625 !
686 !
579
WritePixelLine ! 227,528 !
25,845 ! 21,581
DrawEllipse ! 62.254
! 11,875 !
10,432
DrawCircle !
59,734 ! 12,280
! 11,032
Draw
! 24,210 !
4,447 ! 20,619
Draw Hor/Ver ! 87,029
! 27,791 !
21,281
ScrollRasterX ! 1,755
! 403
! 2,251
ScrollRasterY ! 1,817
! 394
! 2,377
PutText !
60,443 ! 11,085
! 9,708
BlitBitMap !
1,016 ! 10,657
! 15,274
BlitBitMapRP ! 1,013
! 8,688 !
10,963
BitMapScale !
323 !
93 !
94
..........................................................................
OpenWindow !
781 !
156 !
122
MoveWindow !
4,043 !
398 !
531
SizeWindow !
1,094 !
124 !
131
CON-Output !
2,747 !
503 ! 1,227
ScreenToFront[2] ! 51
! 83
! 60
Test (800x600x16)! Amithlon ! 060-50/Mediator/V3-2000
---------------------------------------------------------
RectFill !
7,030 ! 17,794
RectFillPat. ! 1,505
! 16,801
WritePixel ! 1,595,647
! 194,676
WriteChunkyP. ! 3,368
! 288
WritePixelArr. ! 3,367
! 285
WritePixelLine ! 134,264 !
16,158
DrawEllipse ! 58,342
! 11,050
DrawCircle !
55,532 ! 11,610
Draw
! 17,078 !
22,784
Draw Hor/Ver ! 60,072
! 26,385
ScrollRasterX !
516 ! 971
ScrollRasterY !
522 ! 1,002
PutText !
60,595 ! 12,236
BlitBitMap !
597 ! 17,194
BlitBitMapRP !
593 ! 13,037
BitMapScale !
307 !
29
........................................................
OpenWindow !
768 ! 158
MoveWindow !
3,140 ! 893
SizeWindow !
1,053 ! 196
CON-Output !
1,164 ! 1,176
ScreenToFront[2] ! 51
! 60
%%%%%%%%%%%
% RC5/OGR %
%%%%%%%%%%%
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!
Results:
* dnetc client 2.8015-469-CTR-01050611:
Comparison values: <3AFB321E.MD-1.4.4.sgio@erols.com>
Amithlon, 1.3GHz Athlon TB,
RC5: 1.88 MKeys/s
A3000/CSPPC/604e-280,
RC5: 0.93 MKeys/s
A3000/CSPPC/060-66,
RC5: 0.20 MKeys/s
Amithlon, 1.3GHz Athlon TB,
OGR: 3.47 MNodes/s
A3000/CSPPC/604e-280,
OGR: 3.65 MNodes/s
A3000/CSPPC/060-66,
OGR: 0.42 MNodes/s
* dnetc client 2.8010-461-CPR-00062713:
Comparison values: http://n0cgi.distributed.net/speed/index.html
Amithlon, 1.3GHz Athlon TB,
RC5: 2.20 MKeys/s
A3000/CSPPC/604e-200(!),
RC5: 0.66 MKeys/s
A3000/CSPPC/060-66,
RC5: 0.20 MKeys/s
Amithlon, 1.3GHz Athlon TB,
OGR: 3.42 MNodes/s
A3000/CSPPC/604e-200(!),
OGR: 1.60 MNodes/s
A3000/CSPPC/060-66,
OGR: 0.34 MNodes/s
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Benchmarks for which I do not have comparison values available
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Here are some benchmarks which I think are good indications of "real
world
stuff", but for which I, as of now, lack (reliable) comparison values
from real
060 and/or 604e Amigas. If you have such results, please send them
my
way. Thanks!
%%%%%%%%%%%%%%%%%%%%
% Povray/Povbench %
%%%%%%%%%%%%%%%%%%%%
This one is fairly memory bandwidth sensitive, and also might thrash
the
branch prediction cache on some processors.
Instructions: Download a version of Povray compiled for your processor/machine.
Amiga binaries used to be available at
http://www.amigaworld.com/support/povamiga
I currently can't connect to that host, though.
Then issue the following commands:
assign povray31: your_path_to_povray
povray31:
povray31 -i skyvase.pov +v1 +ft -x +mb25 +a0.300
+j1.000 +r3 -q9 -w640
-H480 -S1 -E480 -k0.000
-mv2.0 +b1000 +MB25
[Note: I broke that last line for readability; It's all one line!]
Version used here: Version 3.1 SAS/C-Amiga, Copyright 1998
Result: Amithlon, 1.3GHz Athlon TB --> 104 seconds
%%%%%%%%%
% bzip2 %
%%%%%%%%%
This one is using a near-state-of-the-art file compressor. bzip2 is
a block sorting compression program, using the Burrows Wheeler Transform.
It also does a fair bit of bit fiddling when it comes to entropy-coding
the move-to-front ranks (Yes, I am a compression person. If that was
all
incomprehensible to you, don't worry!).
bzip2 is available on Aminet, at
http://www.aminet.net/pub/aminet/util/arc/bzip2.lha
That archive includes versions for 68k, the Phase5 PPC.library, and
WarpOS; Thus, it makes an ideal comparison testbed.
Once you have the appropriate version for your machine installed (read
the README! And don't forget that you might need to set a sizable stack,
or things *will* crash!), get
http://byron.csse.monash.edu.au/forsteve/oink.bz2.bin
(which is a 150k file, bzip2'ed from 4MB). Bunzip it with any version
you like (this isn't the test yet), and then simply time how long the
following takes, using the version of bzip2 you installed:
bzip2 -9 oink
where "oink" is the name you bunzip2'ed the downloaded file to. This
might
take several minutes, especially on the 060, so don't be impatient.
This is best done in a ramdisk, so your results are not influenced
by
disk I/O
Result: Amithlon 1.3GHz Athlon TB --> 34 seconds
%%%%%%%%%%%%%%
% Voxelspace %
%%%%%%%%%%%%%%
This is a demo that comes with the WarpOS distribution. It can use either
a 68k CPU or a PPC, if available. In both cases, it is executing hand-written
assembler code. The "Freq:" entry in the information box is a frames-per-second
counter.
Some comparison values exist, suggesting around 100fps in 320x200 on
a
CSPPC/CVPPC combo. However, that resolution is so silly, it's not worth
talking about (I *do* have a 21" monitor, after all!).
Instructions: Run the "Run Voxelspace P96" or "Run Voxelspace CyberGfx"
script, whichever gives better results on your particular setup. Select
"640x480" for the first test, and "1280x1024" for the second. Leave
all other demo settings at their default values, and (without touching
the mouse or keyboard) note down the "Freq:" field. Then move around
a bit and check out the range of fps you get while in various spots,
looking into various directions (all sorts of cache issues come into
play
when doing this).
Results:
Amithlon 1.3GHz Athlon TB, GF2 MX-200, 640x480, P96
-> 140fps (130-170fps)
Amithlon 1.3GHz Athlon TB, GF2 MX-200, 1280x1024, P96 ->
31fps ( 20- 35fps)
%%%%%%%%%
% Quake %
%%%%%%%%%
There are many versions of this game for the Amiga. On Amithlon, we
use
Quake68k by Frank Wille and Steffen Haeuser, available from
http://www.aminet.net/pub/aminet/game/shoot/Quake68k.lha
We also have, currently just for our internal use, a version of this
in
which three texture mapping routines were replaced with x86 native
code.
Doing this took one day (most of which was spent on working out which
routines to target, and where to find suitable x86 replacements for
them).
Frank Wille and Steffen Haeuser's port also comes in various PPC flavours,
and you can also get a GLQuake port which, IIUC, sits on top of Warp3D.
The "standard" Quake speedtest is to go into full screen mode (no
status bar, just plain, rendered stuff) activate the console (the '~'
key will usually do that) and type
timedemo demo2
Then deactivate the console, watch until the demo finishes, reactivate
the console, and there will be an fps number.
Apparently, Amigactive instead uses "timedemo demo1". Oh well....
Here are some results (from the usual machine)
1024x768, pure 68k code, timedemo demo2 : 16.2fps
1024x768, pure 68k code, timedemo demo1 :
16.7fps
640x480, pure 68k code, timedemo demo2 :
34.1fps
640x480, pure 68k code, timedemo demo1 :
34.4fps
1024x768, some x86 code, timedemo demo2 : 22.2fps
1024x768, some x86 code, timedemo demo1 :
23.8fps
640x480, some x86 code, timedemo demo2 :
42.6fps
640x480, some x86 code, timedemo demo1 :
44.1fps
[1] What happens is that the raw, measured value x is taken and divided
by 1. If the result is <5, then it gets entered
into the record.
Otherwise, x is divided by two. If that makes the
value <5, then
it is used. Otherwise, division by 3, 4, 5 and so
on is performed,
until finally a value <5 is achieved.
[2] This pretty much is the same as the screen refresh rate as perceived
by AmigaOS. Unless the value is <50, it has nothing
to do with actual
speed.
%%%%%%%%%%
% Benoit %
%%%%%%%%%%
This is a little fractal renderer that Phase5 included with their CSPPC
cards. It can calculate using either the PPC or the 68k CPU; As Phase5
also distributed source (although not to the binary they included....),
it is quite evident that there are close to no 68k->PPC->68k context
switches when using the PPC to calculate.
The same program is also available from Aminet at
http://www.aminet.net/pub/aminet/biz/p5/Benoit-V2_0.lha
Instructions: Set your screenmode to 1024x768x24 or 1024x768x32 (this
is
*very* important, because the size of the Benoit window is calculated
as
1/4 the size of the current screensize. If you use different screen
sizes,
you cannot compare results!).
Then start the Benoit executable; It will self-time how long it takes
for
the startup fractal. You can then change the CPU settings and click
"Calculate" again, to have the same fractal rendered by your other
CPU.
Results: (usual machine)
Startup Fractal, pure 68k code
: 0.30 seconds
We also have a version, currently for internal use only, in which we
compiled
the actual calculation engine to x86 code.
Startup Fractal, x86 calculations :
0.06 seconds
[ Note: Just recompiling the C source to 68k code with gcc reduced the
68k time to 0.20 seconds, though, so the speedup from going
gcc-compiled
x86 is not quite as massive as it looks ]
BACK
|
