dac@prolix.ccadfa.oz.au (Andrew Clayton) (01/28/91)
In article <1991Jan27.074838.10517@watdragon.waterloo.edu>, Colin Plumb writes: > > In article <NAPO.91Jan12233534@elektra.hut.fi>, Hannu Napari writes: > > > >> Maximum of 1024 iterations is far too little. > > ccadfa.cc.adfa.oz.au!prolix!Dac wrote: > > > >Picky picky. :-). > > > >I find that 1024 is just too slow (and that's on a 30Mhz 68030/68882). > > > >Just how deep into the Mandelbrot set are you going anyway? Past a certain > >point, it's all self referential and derivative anyway! Anything at 256 > >iterations is cool enough for moi! > > Wimp. It is a fact of life that, however much processing speed you have, > you're always going to start generating pictures that take over 20 minutes. Is this some 'aha' experience speaking, or a general truism that you've decided to espouse without using anything firm like facts? :-) > Of course, the last Mandelbrot demo I wrote was for a 28-processor transputer > system, so I had it in places where 15,000 iterations was too fuzzy; I had > to go to 17,000. At 513x513 resolution, this wrapped the flops counter > past 2^32. Right. (512^2) * 17000 = 4.4*10^12 flops. Therefore you are saying that you were doing a screen a second? (what's 'a' flops, hmm?). A 4Gflops machine is very bloody fast. > I had to switch to a floating point flops accumulator! > (Actually, I counted z^2+c iterations and scaled the number of seconds > to produce statistics. But 16,000 times a quarter of a million is 4 billion: > wrap!) > > Really, there are some great spots I couldn't find with lower iteration > levels around n=5,000. Madness. Sheer madness. The TV special 'Chaos' had some pretty spiffy Mandebrot animations, that looked real-time (just going down seahorse valley, seemingly forever). I did n=4000 on MandFXP, and it just took AGES (this was in a window about 50 pixels by 20 pixels!). Tranputerising it would be kinda neat to see, but wasting one's life (and CPU) on looking at the Mandelbrot set, seems kind of a waste to me. C'est la vie. > -Colin Dac -- _l _ _ // Andrew Clayton. Canberra, Australia. I Post . (_](_l(_ \X/ ccadfa.cc.adfa.oz.au!prolix!dac . . I am. -------- I cannot send or receive email. Not to anyone at all. Not even you.
ccplumb@rose.uwaterloo.ca (Colin Plumb) (01/29/91)
ccadfa.cc.adfa.oz.au!prolix!dac@munnari.OZ.AU wrote: >In article <1991Jan27.074838.10517@watdragon.waterloo.edu>, Colin Plumb writes: >> Wimp. It is a fact of life that, however much processing speed you have, >> you're always going to start generating pictures that take over 20 minutes. >Is this some 'aha' experience speaking, or a general truism that you've decided >to espouse without using anything firm like facts? :-) It was something we concluded, based on a magazine article somewhere (we were looking up neat mandelbrot spots) that asserted that the comfortable computation time was one coffee break. It's an observation based on playing with mandelbrot exploration programs. >> Of course, the last Mandelbrot demo I wrote was for a 28-processor transputer >> system, so I had it in places where 15,000 iterations was too fuzzy; I had >> to go to 17,000. At 513x513 resolution, this wrapped the flops counter >> past 2^32. > >Right. (512^2) * 17000 = 4.4*10^12 flops. > >Therefore you are saying that you were doing a screen a second? (what's 'a' >flops, hmm?). A 4Gflops machine is very bloody fast. No, by "flops" I meant "floating point operations", not per second. 16,000 iterations (2^14) average times slightly over 2^18 pixels produces 2^32 z^2+c iterations total (which was what I was actually counting). I used this figure to print an average computation speed for the picture. I remember getting 4 million mandelbrot iterations a second, so we're talking 1,000 seconds, which equals 16 2/3 minutes. We actually got up to half an hour or so. >> Really, there are some great spots I couldn't find with lower iteration >> levels around n=5,000. > >Madness. Sheer madness. The TV special 'Chaos' had some pretty spiffy >Mandebrot animations, that looked real-time (just going down seahorse valley, >seemingly forever). I did n=4000 on MandFXP, and it just took AGES (this was >in a window about 50 pixels by 20 pixels!). I was just trying to illustrate that it's *never* fast enough. I'd like to get a bunch of the BIT 100 MFLOPS floating-point ALU's and build a hardware z^2+c toy. It would cost a few thousand dollars, but give you 100 million Mandelbrot iterations a second. But I'd still manage to run out... -- -Colin
peterk@cbmger.UUCP (Peter Kittel GERMANY) (01/30/91)
In article <1991Jan29.133659.20006@zorch.SF-Bay.ORG> xanthian@zorch.SF-Bay.ORG (Kent Paul Dolan) writes: > >>> Of course, the last Mandelbrot demo I wrote was for a 28-processor >>> transputer system, so I had it in places where 15,000 iterations was >>> too fuzzy; I had to go to 17,000. At 513x513 resolution, this wrapped >>> the flops counter past 2^32. > >> Right. (512^2) * 17000 = 4.4*10^12 flops. > >Much too slow. If you can't do full screen resolution (preferably >1280x1024) to the needed number of iterations (here 17,000, but pick 32K >for a nice number), at the frame rate of the display (say 60 frames per >second) then you don't have enough horsepower yet. Yes, that's the point. I REALLY WANT a flight simulator program where I can fly with mouse control through the valleys of MandelMountains! As long as our computers don't provide this, they are still much too slow! -- Best regards, Dr. Peter Kittel // E-Mail to \\ Only my personal opinions... Commodore Frankfurt, Germany \X/ {uunet|pyramid|rutgers}!cbmvax!cbmger!peterk
dac@prolix.ccadfa.oz.au (Andrew Clayton) (01/31/91)
In article <1991Jan29.133659.20006@zorch.SF-Bay.ORG>, Kent Paul Dolan writes: > >> Wimp. It is a fact of life that, however much processing speed you > >> have, you're always going to start generating pictures that take over > >> 20 minutes. > > > Is this some 'aha' experience speaking, or a general truism that > > you've decided to espouse without using anything firm like facts? :-) > > This is the standard answer a graphics freak makes to the Luddite who tells > him he can't _possibly_ make use of any more computer power. Gee, ta, Kent. So I'm now dethroned to the rank of 'luddite'. :-) I was promised a version of a Mandelbrot program that ran on an Amdahl mainfram (IBM 3090 clone), but the guy who wrote it never released the code to me. Dammit. The pictures he was producing, even on a terrible GDDM graphics terminal, were pretty good. Even took the mainframe a deal of time to get down low, and actually broke after 5 zooms. > > Right. (512^2) * 17000 = 4.4*10^12 flops. > > Therefore you are saying that you were doing a screen a second? > > (what's 'a' flops, hmm?). A 4Gflops machine is very bloody fast. > Much too slow. If you can't do full screen resolution (preferably > 1280x1024) to the needed number of iterations (here 17,000, but pick 32K > for a nice number), at the frame rate of the display (say 60 frames per > second) then you don't have enough horsepower yet. 60FPS is overkill. You can get by with 24FPS (I guess a factor of two isn't that much to want - 1280*1024*32K*60 = 2.576*10^12 = 2,576,000,000,000 operations per second. Which is about 256 times quicker than Control Data's top of the line ETA-10 machine was capable of (10,000,000,000 FLOPS). Might take a while to acheive. :-) > > Mandebrot animations, that looked real-time (just going down seahorse > > valley, seemingly forever). > More likely took weeks to create; the universe does not yet contain > sufficent accessible computing power. -the universe- is a little overstating the point, Kent. > > I did n=4000 on MandFXP, and it just took AGES (this was in a window > > about 50 pixels by 20 pixels!). > Well, I went in to a square 1.0E-13 on a side just zooming into interesting > windows found in previous windows, and it tended to take over a day per > frame, but I can testify there's still interesting new stuff down there. > Being heavily into pain, I did this on an Apple ][+. Yuk. I climbed from the mire of Apple ][+ into Amiga, supposedly to play with graphics stuff. I even transcribed a C program to generate Mandelbrots, that took 45 minutes to do the base Mandelbrot. Sigh. Even with a 68030, it's nowhere near quick enough. (And after talking about thousand gigaflop machines, my humble 8.9Million instructions per second '30 is very weak). I went down to 1/128millionth of the screen, (lots of powers of two), saving each miniscule (1" square) picture that I created, and animating the final result. The animation "went too quick", because 'doubling' loses too much definition. I pested ASDG (Agents for CygnusSoft) to see if the guys who wrote MandFXP3.0 had released a 'movie maker' version. I got a 'no' response. :-( > > Tranputerising it would be kinda neat to see, but wasting one's life > > (and CPU) on looking at the Mandelbrot set, seems kind of a waste to > > me. C'est la vie. > The cost in electricity is small, and if you have the cpu power, and an > eye for a pleasing palette, it can be a georgeous exploration. Too repetitive. Self referential. Candy floss for the mind. > Kent, the man from xanth. Dac -- _l _ _ // Andrew Clayton. Canberra, Australia. I Post . (_](_l(_ \X/ ccadfa.cc.adfa.oz.au!prolix!dac@munnari.OZ.AU . . I am. --------------Phone +61 6 285 2537 (+10GMT) // I cannot currently send email.
cg@ami-cg.UUCP (Chris Gray) (02/01/91)
In article <804@cbmger.UUCP> peterk@cbmger.UUCP (Peter Kittel GERMANY) writes: >Yes, that's the point. I REALLY WANT a flight simulator program where >I can fly with mouse control through the valleys of MandelMountains! >As long as our computers don't provide this, they are still much too >slow! I agree! In full 3-D perspective too. Before it went under due to lack of guts on the part of investors and government, I worked at Myrias Research Corporation, which was building large-scale parallel processors out of off-the-shelf microprocessors. Our latest version, the SPS-3, used the 68040 as an engine, with up to 16M per processor of memory, and fast inter- processor communication. Of course we had a Mandelbrot program running (it's trivial to parallelize). Our problem in the past had been I/O - we had to go through a pipe on UNIX to a program which drew on a workstation's screen, but with our new Input/Output Processors (my babies!) we could go quickly to a frame-buffer. I was SO looking forward to trying out 128 or more '040s making pretty pictures under mouse control! I believe we did run it on 1024 68020's, but the kludgy I/O killed it. -- Chris Gray usenet: alberta!ami-cg!cg CIS: 74007,1165
skipper@motaus.sps.mot.com (Skipper Smith) (02/01/91)
In article <804@cbmger.UUCP> peterk@cbmger.UUCP (Peter Kittel GERMANY) writes: >In article <1991Jan29.133659.20006@zorch.SF-Bay.ORG> xanthian@zorch.SF-Bay.ORG (Kent Paul Dolan) writes: >> >>>> Of course, the last Mandelbrot demo I wrote was for a 28-processor >>>> transputer system, so I had it in places where 15,000 iterations was >>>> too fuzzy; I had to go to 17,000. At 513x513 resolution, this wrapped >>>> the flops counter past 2^32. >> >>> Right. (512^2) * 17000 = 4.4*10^12 flops. >> >>Much too slow. If you can't do full screen resolution (preferably >>1280x1024) to the needed number of iterations (here 17,000, but pick 32K >>for a nice number), at the frame rate of the display (say 60 frames per >>second) then you don't have enough horsepower yet. > >Yes, that's the point. I REALLY WANT a flight simulator program where >I can fly with mouse control through the valleys of MandelMountains! >As long as our computers don't provide this, they are still much too >slow! > >-- >Best regards, Dr. Peter Kittel // E-Mail to \\ Only my personal opinions... >Commodore Frankfurt, Germany \X/ {uunet|pyramid|rutgers}!cbmvax!cbmger!peterk Weeeelllll, what you could do is build a 68040/dual 96002 board for the A3000. The 68040 would handle all of the integer operations that needed to be done, the 96002s each do about 40 MFLOPS, and you relegate the 68030 (the one on the motherboard that is still active on the A3000, remember) to a DMA processor to keep feeding you data. I have heard of someone who is working on just such a product, however I do not believe that it will be commercial (I think it may be for the DoD, but I really don't know) and so I will not mention who they are. However, it would be a particularly WONDERFUL thing for anybody who could afford it! Just out of curiosity, how fast do ya'll think this would crunch data. I think the target of the previously mentioned product was to be able to crunch U.S. Geological Survey data into fractal landscapes in real time (basically, take the VISTA videotape and make it interactive for a basic idea). I don't know if they will make it, but it is an admirable task. -- Skipper Smith | skipper@motaus.sps.mot.com Motorola Technical Training | 8945 Guilford Rd Ste 145 All opinions are my own, not my employers | Columbia, MD 21046