NVIDIA brings supercomputing to the desktop

Wonder if they're capable of running Crysis at a constant 60FPS at maximum detail... :p

I must say I find it fascinating using GPUs for such computations, especially with the likes of Folding@Home :)
 
GPUs are made for graphics and while the can do some stuff in parallel at very high speeds they are often not as accurate in floating point operations as they should be. This is not a problem if it is merely processing pixels that fly past you at a rate of 100 million per second, but it is a problem when you are calculating scientific results.

Distributed computing clients like BOINC do not run on GPUs because of their inaccuracy. Besides that if you look at AMDs push to integrate the GPU on the CPU I really do not see the big hype about GPUs. Soon they will make out one or two of the many cores we find on our CPUs.
 
Rouxenator said:
Distributed computing clients like BOINC do not run on GPUs because of their inaccuracy. Besides that if you look at AMDs push to integrate the GPU on the CPU I really do not see the big hype about GPUs. Soon they will make out one or two of the many cores we find on our CPUs.
Click to expand...

I believe you are incorrect, Folding@Home runs on the new 8 and 9 series GPU's and is probably one of the projects that needs extremely high accuracy.
 
If AMD merged its GPU's with its CPU's then wouldn't it make less money? Seeing as its selling only one product... instead of two?
 
JUGGY said:
I believe you are incorrect, Folding@Home runs on the new 8 and 9 series GPU's and is probably one of the projects that needs extremely high accuracy.
Click to expand...
Folding@Home does not need as accurate computations as some of the projects you can attached to with BOINC.
inferno said:
If AMD merged its GPU's with its CPU's then wouldn't it make less money? Seeing as its selling only one product... instead of two?
Click to expand...
It will be a better product than the competition so more people would buy it. It makes for simpler and cheaper motherboards. It provides a central point of upgrade. It will be better and intel will copy it just as they did with x64 and integrated memory controllers.
 
well... as Rouxenator said, gpu's are inaccurate... they good at assembling geometric shapes, and at placing pixels, but the way they process is somewhat all over the show, you might end up having lost bits of information and such... but gpu's are designed to do precision processing, they aren't designed at calculation proteins, rather, they designed for graphics. and now companies are seeing that gpu's are fast enough to do some general processing. So, when we start to find physics in games become very complex, errors in the processing will become present...

this is why intel is pushing ray-tracing on the cpu, as you might find errors, anomalies and miss calculations with the gpu, and, yes, ray-tracing can be done of the gpu, but for it to be accurate, the gpu will need to double check everything, which means a mega performance hit...

as for the GPU on a CPU post,
you must remember that once AMD incorporates these two processors on one ship, it would mean that the performance increase each year would be greater as they can have increased cpu and gpu performance and people would feel obliged to upgrade more often... Also, if they pull this move, more people would want to buy an AMD cpu/gpu as it's cheaper and possibly better performing then intel...
 
No.
There's nothing inaccurate about a GPU. Precession and Accuracy are not the same thing. GPU is a RISC ASIC and is as accurate as any other processor for the same precision. Single and double precision. GPUs support 128-bit FP32 calculations and FP64 precision that's IEEE 754 compliant. CPU's will not support some of the precision targets that modern GPUs have.

What you may be talking about is instruction set and that's something entirely different. Cg+ is basically C++ for the GPU looks very similar, and behaves in a similar way. The GPU is by far more flexible than CPU these days, it's a massive parallel processor that's very good at threaded programs where as the CPU is the other way 'round.

Ray-tracing has very little if anything to do with accuracy in this context. Ray-tracing is not applicable here as GPUs are raster based renderers and the pipeline is set up accordingly. Vertex in - pixel out. Ray tracing is very good at simulating light interactions as it works at a photon level. However it is horridly slow, cannot do AA and is just not feasible for any real time rendering. Note that CG movies these days are 90% raster based and only 10% if not less traced. Raster based rendering can be as good and some times more realistic than ray-tracing.

but gpu's are designed to do precision processing, they aren't designed at calculation proteins, rather, they designed for graphics. and now companies are seeing that gpu's are fast enough to do some general processing. So, when we start to find physics in games become very complex, errors in the processing will become present...
Click to expand...
One of the key applications for NVIDIA's Tesla system is protein synthesis and physics simulations.
 
Last edited:
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X