I’m looking to get my first GPUs for scientific computing on Macs. I’ll probably first get some Matlab code accelerated using Jacket or GPUlib for the scientists I work with before I handcode with CUDA.
I need double-precision, so am looking at GTX 260/285/295, C1060 and FX 5800 GPUs.
The question is, what really is the difference in computational performance? The GTX 285, C1060 and FX 5800 all seem to be rated at about 1 TFLOPS and have 30 multiprocs. They have different memory. If I’m not worried about memory and the bandwidth between the CPU and GPU (I’ll be loading big data only during one-time initialization), will their performance be different? Will the GTX 295 really be about twice as fast, assuming I split the load properly between the two GPUs?
Why is the C1060 so 2x $ more than the GTX 295 - just memory? And the FX 5800 is so much more than both those, with the same memory as the C1060…why?
I’m assuming I can use all these on Mac Pros if I have the slot open and enough power supply? How about a PowerMac G5 with PCI? If I go for a C1060, would I have the problem I read about with Win XP where you also need either a nvidia card or motherboard chipset that has video?
[*]From a peak computational perspective, all of the 240 core GTX200 basic GPUs are pretty much equal
[*]The C1060 is “industrial strength”. I am pretty sure NVIDIA are harvesting GPUs and applying varying levels of test and QC limits. The Telsas are engineered for optimal reliability in production environments. The less expensive consumer video cards less so.
[*]You can’t use anything other than the Mac specific version of the GTX285 in an Intel Mac. The Mac requires a card with EFI firmware rather than the old fashion PC VGA video BIOS. Power macs are not supported at all.