Slashdot Mirror
Future of 3d Graphics
zymano writes "Extremetech
has this nice article on the future of 3d graphics. The article also mentions that graphic card gpus can be used for non-traditional powerful processing like physics. A quote from the article, "GPU can be from 10 to 100 times faster than a Pentium 4 and Scientific computations such as linear algebra, Fast Fourier Transforms, and partial differential equations can benefit". My question - If these cards are getting so powerful at computations then why do we need a Intel/AMD processor at all? Just make a graphics card with more transistors and drop the traditional processor..."
My question - If these cards are getting so powerful at computations then why do we need a Intel/AMD processor at all? Just make a graphics card with more transistors and drop the traditional processor...
If you'd really like the answer to this question, try programming anything on the GPU and you'll understand. It's hell to do half this stuff. GPUs are highly specialized and make very specific tradeoffs in favor of graphics processing. Of course, some operations, specifically those that can be modeled using cellular automata, map well to this set of constraints. Others, such as ray-tracing can be shoe-horned in, but if you were to try to write a word processor on the GPU, it'd essentially be impossible. The GPU allows you to do massively parallel computations, but penalizes you heavilly for things such as loops of variable length or reading memory back from the card outside of the once-per-cycle frame update, and the price of interrupting computation is prohibitive. Clearing the graphics pipeline can take a long, long time.
Furthermore, while there have been a few papers published claiming the orders of magnitude increase in speed in these sorts of computations, none actually demonstrate this sort of speed-up. Everyone's speculating, but when it comes to it, results are lacking.
b.c
Actually, I see it more like having a brain surgeon as your family doctor.
No comment.
While optimized for graphics, GPUs can indeed be used as general-purpose processors. GPUs are effectively stream processors, a class of devices whose architecture and programming model make then particularly efficient for scientific calculation.
> It might take a real long time, but it is a general purpose processor and so can process anything
The same holds true for GPUs. Like CPUs, they are turing complete.
Not true. Newer cards appear to be IEEE 'extended' (there isn't a definition of the bitsize of extended from the spec, so even Intel's 80-bit format is considered 'extended') at 128 bits wide per color channel. This is pretty much the last word in accuracy as far as I'm concerned. Perhaps numerical analysts can come up with scenarios where 128 bits aren't sufficient, but I don't want to hear about them.
For some of the stuff that we do, we would kill for a slightly faster card. Right now, for simulation of IR imagery, we have to prefly a scenario where the sensor-carrying vehicle (use your imagination) flys a trajectory and we render the imagery along this path. This rendering consists of doing convolutions of background scenes with target information to generate a final image. At the end we have a 'movie'. This can take a few hours to run.
Afterwards, we run the simulation in realtime and play frames from this movie (adjusted in rotation and scaling, etc. because real-time interactions can result in flight paths subtly different from the movie) and show it to a *real* sensor and see what happens.
The point: if we could do real time convolution inside a graphics card and then get the data back out some way (we usually need to go through some custom interface to present the data to the sensor), then a lot of pain would be saved. First, we could move the video-generating infrastructure into the real time simulation, which would be simpler, we wouldn't have to worry about rotating and scaling the result since we'd be generating exactly correct results in the fly, we wouldn't have to worry about allocating huge amounts of memory (Gigabytes) to hold the video and all the concerns about memory latency and bandwidth and problems with NUMA architectures, and finally (maybe) we could change scenarios on the fly without having to worry about whether we already had a video ready to use.
I think the computational horsepower is almost there, but right now there's no good way to get the data back out of the card. On something like an SGI you get stuff after it's gone through the DACs, which mean you now have at most 12-bits per channel (less than we want, although you can use tricks for some stuff to get up to maybe 16-bits for pure luminosity data). What would be sweet in the extreme is to get a 128-bit floating point value of each pixel in the X*Y pixel scene. So if the scene were 640x480 then we'd get about 4.5Meg of data per frame at say 60Hz then we'd get about 281Meg a second to convert and send out.
Life would be sweet. Sadly, this is a pretty special purpose application, so I'm not too hopeful. What's weird is that only NVidia (and perhaps ATI) are coming up with this horsepower because of all the world's gamers, and vendors like SGI are left with hardware that is many, many generations old (although it does have the benefit of assloads of texture memory).
In short: need 1GB of RAM on the card and a way to get stuff back out after we've done the swoopty math.
OpenGL and Direct3D are the two interfaces that graphics card vendors provide to get at the hardware; there is no lower-level way to get at it. However, these APIs now include ways of describing programs that run on the GPUs directly; you can write programs that run at the per-vertex or per-pixel level with either of those APIs.
These programs can be given to the GPU via specialized low-level assembly language that has been developed to expose the programmability of GPUs. (They are pretty clean, RISC-like instruction sets).
Alternatively, you can use a higher level programming language, like NVIDIA's Cg, or Microsoft's HLSL, to write programs to run on the GPU. These are somewhat C-like languages that then compile down to those GPU assembly instruction sets.