bandwidth question

Accelerated Computing CUDA CUDA Programming and Performance
1

I use GTX 280, chipset Intel X38, CPU Intel Core2Quad, Red Hat Enterprise Linux Server release 5.2, CUDA toolkit 2.0, sdk 2.0
Motherboard supports PCI e 2.0 x16 , so as I understand peak one-way bandwidth must be 8 GB/s
But after running bandwidth test from sdk I got these results:

Pageble:
Host2Device 2419.1 MB/s
Device2Host 2343.3 MB/s
Device2Device 115377.7 MB/s
Pinned:
Host2Device 2521.7 MB/s
Device2Host 3274.7 MB/s
Device2Device 115356.9 MB/s

3.3 GB/s is less then a half from peak bandwidth!!! Are my test results OK? If not, what could be the problem?

2

I just built a new GPU server using a completely different motherboard and CPU (Phenom + AMD 790FX chipset), but was surprised to find that out of the box, the motherboard disabled PCI-e 2.0. The results from the bandwidth test looked very similar to yours.

To fix it, I had to go find the option instructing the BIOS to auto-detect the PCI-Express 2.0 capability of the card at boot, rather than assume 1.0.

Maybe you are seeing a similar problem.

3

Thanks for advice, I ll try that.
After searching the forum, I found that the solution could be updating BIOS to the newest version, maybe this will help.