Using C870 with NVS 285 Performance Bad

Accelerated Computing CUDA CUDA Programming and Performance
1

I’m trying to use Tesla C870 with NVS285 x1.

But when I run samples from SDK 1.0.

Performance is very bad.

ex.histogram64
GPU time is 58XXmesc // 16.xxMB/sec
CPU tims is 544msec // 171.xxMB/sec

I don’t know how could this happen.

GPU is worse than CPU. >.<

This is my environment
MB:Tyan S2895 nVIDIA nForce Pro 2200 + 2050
CPU:2x DualCore AMD Opteron 285, 2600 MHz
MEMORY:A-Data 2 GB PC3200 ECC DDR SDRAM * 8
OS:XPP SP2
C870 in slot 1
NVS285 x1 in slot 3
Both slots are x16.
DRIVER:162.61 (or 162.65 in the TESLA disc)
CUDA toolkit & SDK:1.0 (in the disc)

Are there any possible that I’m running CUDA by 285?

Does any using C870 with other card too?

Thanks…

2

Are you sure you’re not running emulation builds?

3

YES

You’re right.

I was running files under EmuRelease.

It works well, but I can’t run anything about “GL” on C870.

If I use 4600 than I can run “GL”.

THX

4

Multiple GPU configurations with OpenGL interop do not work in CUDA 1.0. You will need to use CUDA 1.1 for that. Note, though, that CUDA 1.1 solves the problem by transferring the data from one card to the other through system ram and may be much slower than running both computation and display on the same device.