Is it possible to us GeForce 9800 GX2 on a server motherboard without SLI for CUDA?
Can I us two cards (4 GPU) on Intel 5400 motherboard with 2xPCI Express Gen2 x16?
Thank you! :)
Yes and probably yes.
Make sure you have at least one CPU core per GPU, else you will get a severe performance penalty.
Just so I wont get bitten when moving from small scale performance testing to full scale testing. Is this also the case when you are using streams? As far as I understand, you can query whether a stream is finished, without blocking. So using streams & contexts you should be able (as far as I see it) to keep more than 1 GPU busy from 1 core.
I think the main issue will be latency if you have short kernels. A hot spin on the CPU side will let you catch the termination of the GPU kernel very quickly, which could be important if you are rapidly alternating between CPU and GPU code. If your kernels take a long time to run, then it probably won’t be so bad.
(Edit: Hah, looks like this was already discussed in a related thread. Please excuse my redundancy.)
We (rZrzucek and me) would like to buy a platform for research and development of real-time algorithms. The system will work with video stream (15fps, max 30fps) that is why, the output data will be calculated every one frame (1-5 frames delay is acceptable). The algorithm is highly parallel, we would like to use CUDA in order to accelerate part of the calculation and OpenGL to 3D visualization in the same time on the same computer.
Which platform would you like to use for such a problem?
4 x Opteron (2.2GHz, 2 core each) + 16GB RAM
2 x Xeon (~2.4GHz, 4 core each) + 16GB RAM
System with Linux OS, GeForce 9800 GX2, and probably TESLA C870. Another question: Is it possible to use GeForce 9800 GX2 and TESLA accelerator in the same time for different purpose (OpenGL acceleration on GeForce and CUDA on TESLA) on condition we have motherboard with 2x pci-e x16 (eg. UniWorkstation 5548ES 4way AMD Opteron , 2xNvidia SLI 1000W) ?
You can use a 9800 and a tesla in the same machine yes. Just be aware that the data will be copied from one card to the other (slow) for display. If close to realtime display is what you are looking for, your best bet is to do display and compute on the same card. Though, you’ll have the same problem with the 9800 GX2 because it is really 2 cards in one.
For a GPU, I would recommend a 8800 GTX, Ultra, or equivalent Quadro (if more memory is needed). Your video processing algorithms are likely to be memory bound, not compute bound and these cards have faster memory interfaces than the 9800 GX2.
For a CPU, only you know what you really need. But I think 8 cores is way over the top. Typicaly CUDA accelerated performance is ~30x faster than a single core so anything that goes back to the CPU just slows things down. That and you only need 1 CPU core per GPU.