Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Extra Processing Power - Cheap?
- Thread starter Michael73
- Start date
- Sort by reaction score
You're using software that can actually use it and it's extended driver capabilities?
To go back to the original question....
I am retro-fitting 08 Pros with GTX 200-series cards, initially for use under BootCamp but now hacking for OS X while I wait for the Mac 285. I will be spending my money here rather than on 09 Pros. This is all for CUDA/OpenCL work, and to be quite honest it is an experiment to find out the performance benefits that you want to know about as well.
It is probably too early to answer your question properly. When we get Snow Leopard, proper support for a variety of GPUs and lots of tools like routine OpenCL acceleration of photo, video and math it will be clearer to jduge the benefits. But we do know some single precision computations can easily go over two orders of mag faster, double precision (right now) at least one. Something doing a lot of writes to the disk will not benefit so much as the bottleneck is elsewhere. Apart from the CPUs, there does not seem to be a huge difference between the 08 and 09 architectures - both have two PCI slots with the right interfaces to run two GPUs, though the power routing needs some fiddling.
http://www.nvidia.com/object/cuda_home.html#state=home
has a few nice examples of speedups. There is a link to a paper showing the solution of stochastic differential equations (single precision) going 675 times faster than a CPU when done on a TESLA C1060 compared to a single core on 2.66GHz Core2Duo.
For CUDA/OpenCL work I am trying to figure out where all these GPU cards are on the price-performance curve. My current test rig has an 8800 and a 285, the latter with 2Gig and 240 cores, and it runs at just under 1.5GHz as a CUDA engine, for < 325GBP. As a CUDA engine, a Quadro 4800 has fewer cores, less memory (and possibly lower clock speed) for 4 times the money. If someone shows the 285 produces actual math errors **under CUDA/OCL** I will rethink my preference. This is rather critical to the price/perf question and I do not know the answer, though I have yet to find math errors on my cards so far. If this remains the case the Quadro is probably irrelevant as a general GPU accelerator.
I am retro-fitting 08 Pros with GTX 200-series cards, initially for use under BootCamp but now hacking for OS X while I wait for the Mac 285. I will be spending my money here rather than on 09 Pros. This is all for CUDA/OpenCL work, and to be quite honest it is an experiment to find out the performance benefits that you want to know about as well.
It is probably too early to answer your question properly. When we get Snow Leopard, proper support for a variety of GPUs and lots of tools like routine OpenCL acceleration of photo, video and math it will be clearer to jduge the benefits. But we do know some single precision computations can easily go over two orders of mag faster, double precision (right now) at least one. Something doing a lot of writes to the disk will not benefit so much as the bottleneck is elsewhere. Apart from the CPUs, there does not seem to be a huge difference between the 08 and 09 architectures - both have two PCI slots with the right interfaces to run two GPUs, though the power routing needs some fiddling.
http://www.nvidia.com/object/cuda_home.html#state=home
has a few nice examples of speedups. There is a link to a paper showing the solution of stochastic differential equations (single precision) going 675 times faster than a CPU when done on a TESLA C1060 compared to a single core on 2.66GHz Core2Duo.
For CUDA/OpenCL work I am trying to figure out where all these GPU cards are on the price-performance curve. My current test rig has an 8800 and a 285, the latter with 2Gig and 240 cores, and it runs at just under 1.5GHz as a CUDA engine, for < 325GBP. As a CUDA engine, a Quadro 4800 has fewer cores, less memory (and possibly lower clock speed) for 4 times the money. If someone shows the 285 produces actual math errors **under CUDA/OCL** I will rethink my preference. This is rather critical to the price/perf question and I do not know the answer, though I have yet to find math errors on my cards so far. If this remains the case the Quadro is probably irrelevant as a general GPU accelerator.