My current GPU is of CC1.1 and I cannot perform double precision calculations.
It seems like all the CC1.3 GPUs come on a PCI-E 2.0 board but my motherboard is 1.0
Will a PCI-E 2.0 board work on a PCI-E 1.0 motherboard; will it have any effect on CUDA computing, other than memory bandwidth ?
Thanks in advance.
Yes, compute 1.3 PCI-e 2.0 parts work fine in PCI-e 1.0 motherboards, albeit with roughly half the host-device bandwidth.
Make sure there are no obstructions on the motherboard to a long double-slot card. I found that motherboards old enough to only have PCI-Express 1.0 often have a layout that is not particularly well-designed for large cards. I’d discover stuff like capacitors and heatsinks that stuck up too high, or important ports on the board that were covered by the card.
But other than those physical layout issues, I’ve never had a problem running a PCI-Express 2.0 card in a 1.0 motherboard.
That is a good point. South bridge and SATA port placement are the usual ones that can be problematic, especially on micro-atx and smaller server boards.
Thank you both for the input.
And, yes, the heatsink for southbridge and sata ports get in the way.
Looks like mb upgrade is needed.
Do you know of a LGA 775 motherboard with pci-e 2.0 that can handle these type of GPUs?
Regards
What form factor are you after? ATX or micro-ATX?
Well, I got the upgrade bug and will be getting a Zalman tower which will support both (in for a penny in for a pound).
Suggestions for both form factors are greatly appreciated.
By the way the GTX260 card I found is over 13" (~33cm) long, is this typical?
I still can’t believe all this trouble for getting double precision support on a development machine!
Kind Regards
On the ATX front, I don’t have any recent experience, but on the micro-atx front we had quite a lot of recent success with the Intel DQ45CB, which is a pretty straightforward, unexciting legacy-free (no IDE, not floppy, no PS/2 connectors) board which takes a single dual slot full length GPU without much trouble. Tthe front panel connectors sit in the board path, but they don’t interfere with anything. PCI-e gives over 5GB up/down with a Q9650 and 800MHz DDR2 ram,
That’s huge. The Tesla cards are only 10.5" and I think the GTX 280 / 285 are about the same size.
I just wanted to thank everyone who tried and helped me with this issue.
I am glad to say that I could get my current pci-e 1.0 mb to accept the GTX260^2 card.
I lost the ability of using one of the sata ports, but for me that was not a reason enough to upgrade a mb (and endure the rebuilding of my environment).
Other than that, I had to file off couple of capacitors (just kidding).
The card dimensions (13" in length) I was concerned about were not an issue, the manufacturer was probably quoting the box dimensions!
The GPU is a GTX260^2 not 260 and cudaGetDeviceProperties() is reporting 27 multiprocessors, not 24. I could not find a reference to this device in the CUDA Programming Manual 2.2.1.
Best Regards and Happy Holidays to everyone
Earlier GTX260 cards had 192 cores, the later version had 216 core. You clearly have the latter.