Calculations on the GPU?
fungus
> Hi,
>
> I would like to try to write a multispin ising simulation on a 128 or
> 256 bits processor. Would it be possible to use the GPU of a graphics
> card to do the calculations (I need the atomic operations &, |, ^, <<
> and >> )
>
Forget it. Even if it were possible (which it isn't), your
normal CPU is going to be faster for this sort of thing.
--
<\___/>
/ O O \
\_____/ FTB.
Poison
> 256 bits processor.
128 or 256 bits ???
this is a sure mistake
there is no such thing as 128 bit procesor in comercial hardware
even if the GPU has 128 I/O bits it is working in PISO so it won't
achieve the speed you want.
if the speed is not the thing you want, then why do you want
to use such procesors ?
Poison64
pois...@poczta.onet.pl
Poison
it is posible just uninstall it from graphics card and install
it in your circuit :))
(sorry for my english, i have very little knowledge of
technical words)
Poison64
pois...@poczta.onet.pl
White Flame (aka David Holz)
news:an1o3v$3ne$1...@news.onet.pl...
> 128 or 256 bits ???
> this is a sure mistake
> there is no such thing as 128 bit procesor in comercial hardware
> even if the GPU has 128 I/O bits it is working in PISO so it won't
> achieve the speed you want.
um, IIRC PowerPC G4 is fully 128-bit, as are a lot of the newer GPUs (if not
bigger).
fungus
SGI machines have a lot more than that. The old ONYX
machines had 1024 bits per pixel and up to 20 rasterizing
chips going at the same time - that's 20480 separate data
lines in there somwhere (that's why the boxes are so big).
The newer ones have even more bits per pixel, I'm not
quite sure how many as I'm a bit out of touch with SGI
hardware these days but it's definitely "more".
Phil Frisbie, Jr.
>
> ]> I would like to try to write a multispin ising simulation on a 128 or
> a bummer, now I'll have to stick to the Dec :-(
There is a big difference between memory access width and integer math word
width which is what seem to be referring to.
Phil Frisbie, Jr.
Hawk Software
http://www.hawksoft.com
Poison
> machines had 1024 bits per pixel and up to 20 rasterizing
> chips going at the same time - that's 20480 separate data
> lines in there somwhere (that's why the boxes are so big).
>
> The newer ones have even more bits per pixel, I'm not
> quite sure how many as I'm a bit out of touch with SGI
> hardware these days but it's definitely "more".
first of all, it is not a comercial hardware,
and second, "in there somwhere" so you cant use it
externally. Imagine a procesor with 1024 legs per port :))
Poison64
pois...@poczta.onet.pl
Ruud van Gaal
With a 128-bit CPU most people mean a basic register is 128-bits wide.
Are you sure a G4 has 128-bit registers? Can't imagine; Windows is
still a 32-bit OS, and IRIX can just be had at 64-bit.
Probably, on these systems, a C/C++ 'int' is 128-bit wide. Have never
heard of that before, so I'm doubtful about a G4 (I've got an old
350Mhz one, surely doesn't feel like 128-bit) having 128 bits in each
register.
Memory access, sure, but that's a different thing.
Ruud van Gaal
Free car sim: http://www.racer.nl/
Pencil art : http://www.marketgraph.nl/gallery/
Charles E Hardwidge
Small note; Bits may refer to internal registers or external bus, and
required bandwidth != required precision.
--
Charles E Hardwidge
http://www.hardwidge.org.uk
zed
> Hi,
>
> card to do the calculations (I need the atomic operations &, |, ^, <<
> and >> )
>
>
> regards
>
>
ive got no idea what a ising is so i could be talking out my ass but
contary to what others have said.
i use the graphics card for calculations often cause of the great speedups.
eg tiled perlin noise effect of 1/4 million pixels
cpu: 4 seconds
graphics card: 0.05 seconds ie 80x quicker
the 2 things are
A/ u dont have the flexibilty (equation wise etc) of a cpu (though this
will change with new graphics cards, though if they will achieve the
flexibilty of cpus is debatable)
B/ u can't ask for the data back after ech pixel eg in the above example
i do all the calculations + then ask for all the 1/4results in one fell
swoop.
zed
César Blecua Udías
I'm not sure about *total* bits per pixel (ie: including bits for fast
clears, window-IDs, and so on), but if we're talking about color depth,
the newest SGI workstations have exactly the same as older RE/IR beasts
(Onyx/Crimson), which is 12+12+12+12. The difference is that you can
have it on the desk rather than on a big box at the desk side :)
Well, V8 and V12 Fuel and Octane2 support a 96-bit (24+24+24+24)
hardware accumulation buffer, which is (IIRC) deeper than Onyx/Crimson,
but for pbuffers and windows you're limited to 12+12+12+12 (less than
I'd like to have for (really) photorealistic rendering, but it's
obviously better than 8 bits per component).
César
cesar...@ono.com
Andy V
>
> ive got no idea what a ising is so i could be talking out my ass but
> contary to what others have said.
> i use the graphics card for calculations often cause of the great speedups.
> eg tiled perlin noise effect of 1/4 million pixels
> cpu: 4 seconds
> graphics card: 0.05 seconds ie 80x quicker
>
> the 2 things are
> A/ u dont have the flexibilty (equation wise etc) of a cpu (though this
> will change with new graphics cards, though if they will achieve the
> flexibilty of cpus is debatable)
> B/ u can't ask for the data back after ech pixel eg in the above example
> i do all the calculations + then ask for all the 1/4results in one fell
> swoop.
That's a good way to use the GPU. However, if you want data that can't
be stored in pixels in the frame buffer, the use of the GPU becomes
problematic. You don't have general access to the frame buffer memory,
and you might be missing looping controls (I believe the latest
generation
does have loops).
--
Andy V (OpenGL Alpha Geek)
"In order to make progress, one must leave the door to the unknown
ajar."
Richard P. Feynman, quoted by Jagdish Mehra in _The Beat of a Different
Drum_.
Paul Martz's OpenGL FAQ:
http://www.opengl.org/developers/faqs/technical.html
White Flame (aka David Holz)
news:3d96aefd....@news.xs4all.nl...
> With a 128-bit CPU most people mean a basic register is 128-bits wide.
> Are you sure a G4 has 128-bit registers?
AFAIR, it can do full 128-bit arithmetic (straight or SIMD), and its data
bus is 128 bits wide.
White Flame (aka David Holz)
> > The newer ones have even more bits per pixel, I'm not
> > quite sure how many as I'm a bit out of touch with SGI
> > hardware these days but it's definitely "more".
>
> first of all, it is not a comercial hardware,
WTH? Since when have they been non-commercial? :-P
> and second, "in there somwhere" so you cant use it
> externally. Imagine a procesor with 1024 legs per port :))
This simply goes back to what you define a "bit-ness". :)
César Blecua Udías
>
> zed wrote:
> >
> > ive got no idea what a ising is so i could be talking out my ass but
> > contary to what others have said.
> > i use the graphics card for calculations often cause of the great speedups.
> > eg tiled perlin noise effect of 1/4 million pixels
> > cpu: 4 seconds
> > graphics card: 0.05 seconds ie 80x quicker
> >
>
> That's a good way to use the GPU. However, if you want data that can't
> be stored in pixels in the frame buffer, the use of the GPU becomes
> problematic.
I think NVIDIA has announced extensions for floating point RGBA buffers.
Anybody knows more details about this? Will it be hardware accelerated
or software implemented? Having floating point buffers is the golden
dream of developers interested in multipass hardware accelerated
photorealistic rendering.
And, of course, such FP buffers would enable to store more types of data
into pixels, so the gfx hardware can better substitute the CPU for more
tasks...
César
cesar...@ono.com
Ruud van Gaal
<cesar...@ono.com> wrote:
>And, of course, such FP buffers would enable to store more types of data
>into pixels, so the gfx hardware can better substitute the CPU for more
>tasks...
It's weird; we seem to be moving to a world where the CPU is
handicapped, and the GFX card is a mini-computer inside the PC, which
is far beyond the architecture (read: speed/efficiency) of the
computer that it lives in.
A gross re-architecture of the PC design is then in order, I'd say.
But ofcourse, that has been true of the PC for at least 10 years. ;-)
César Blecua Udías
>
> On Sun, 29 Sep 2002 11:35:14 +0200, César Blecua Udías
> <cesar...@ono.com> wrote:
>
> >And, of course, such FP buffers would enable to store more types of data
> >into pixels, so the gfx hardware can better substitute the CPU for more
> >tasks...
>
> It's weird; we seem to be moving to a world where the CPU is
> handicapped, and the GFX card is a mini-computer inside the PC, which
> is far beyond the architecture (read: speed/efficiency) of the
> computer that it lives in.
If the PC graphics card becomes responsible of more and more tasks,
that's good news (IMHO) because that's an evidence that things work best
when the machine is designed as a whole with an unitary concept rather
than as an anarchic mix of differently designed parts. If, for example,
we arrive to an era when the graphics card can run an application with
little or zero CPU work, I'll be happy: No need for PCs anymore: Just a
graphics card, a power supply, and let's go :)
Now I need a graphics card capable of running a C/C++ compiler and a
text editor. Hmmm... I'm waiting you, gfx vendors! :)
César
cesar...@ono.com
Charles E Hardwidge
>tasks, that's good news (IMHO) because that's an evidence
>that things work best when the machine is designed as a whole
>with an unitary concept rather than as an anarchic mix of
>differently designed parts.
Hear, hear, for quality standards.
>If, for example, we arrive to an era when the graphics card can
>run an application with little or zero CPU work, I'll be happy:
>No need for PCs anymore: Just a graphics card, a power
>supply, and let's go :)
I've been seriously thinking this might be the reality in the near future as
waste disposal requirements become more stringent in the European Union and
"developed world." Making so many components redundant, as we do at present,
is wasteful and unnecessary. A high bandwidth streaming architecture with a
plug-in "core" might be the next step. The name "Playstation 3" springs to
mind. Whatever happens, I'd like to see a little more "progress" rather than
"change", which is what we are stuck with at present.
Apologies for the ramble, it's a pet peeve.
zed
>
>>That's a good way to use the GPU. However, if you want data that can't
>>be stored in pixels in the frame buffer, the use of the GPU becomes
>>problematic.
>>
>
> I think NVIDIA has announced extensions for floating point RGBA buffers.
> Anybody knows more details about this? Will it be hardware accelerated
> or software implemented? Having floating point buffers is the golden
> dream of developers interested in multipass hardware accelerated
> photorealistic rendering.
>
> And, of course, such FP buffers would enable to store more types of data
> into pixels, so the gfx hardware can better substitute the CPU for more
> tasks...
>
> César
> cesar...@ono.com
>
a float per R,G,B.
redefine RGB to x,y,z thus turn your screen into a 3dcoordinate system,
which opens up some interesting possibities,
IIRC there are papers on the net they list a few for those that are
interested
zed
dominic glynn
> a float per R,G,B.
> redefine RGB to x,y,z thus turn your screen into a 3dcoordinate system,
> which opens up some interesting possibities,
> IIRC there are papers on the net they list a few for those that are
> interested
definately interested!
can you point me in the right direction to these papers you write of,
(url's, etc) to find some decent resources on achieving this.