5 different compute capabilities in one box

Accelerated Computing CUDA CUDA Programming and Performance
1

[font=“Arial”]I really wanted to simplify my CUDA testing and since I had two spare PCIe x1 slots in my workstation I modified two older cards to fit.

A Dremel tool, safety glasses and a steady hand were required. The low-end cards are now x1 devices.

Now I can test against most compute capabilities without using another machine or swapping cards. No 1.0 or 1.3 but 1.1 and 1.2 are close enough to those compute capabilities.

Here are some bandwidth numbers in MB/sec:

gpu 	pcie    sm_cc  h>d   d>h   device

------- ------- -----  ----  ----  ------

GTX 680 2.0 x8  3.0    2628  3250  155331

GTX 480 2.0 x8  2.0    2638  3277  147911

GT 545  2.0 x8  2.1    2642  3201   35650

GT 240  1.1 x1  1.2 	165   205   21224

9400 GT 1.1 x1  1.1 	165   205    7083

The last two GPUs are in the PCIe 1.1 x1 slots so those numbers aren’t a surprise.

I haven’t smelled any smoke yet! External Image[/font]

2

Great! Can you post the shared memory bandwidth number also? Thanks! External Image

3

One more thing, I just received the GTX 680 and it basically sits at 24C – room temperature – when not in use.

Amazing!

4

Is there a benchmark you would suggest?

The PCIe x1 cards are still running their SM’s at full speed so they’re still useful for testing and benchmarking “GPU-only” kernels.

5

I am not aware of the any benchmark for this. I learned about this value from here:

This number sort of explains why 680 is slower than 580 in SP performance.