dear all:
I have a video card, ASUS ENGTX295, it has two GPUs (each has 896MB RAM) in one card.
I want to merge two GPUs into one, but from NVIDIA control panel, I
cannot see any SLI option.
cudaGetDeviceCount() always reports 2 GPUs, and I can only use 800 MB RAM.
How can I merge 2 GPUs into one such that I can use 1.7GB RAM.
dirver version: 190.38
cuda toolkit and SDK: 2.3
Lung Sheng
As you said, the 295 is two GPUs, not one. There’s no way to make them appear to be a single superGPU in software or even in hardware (they have separate memory busses).
There is indeed 1.6GB of RAM, but it’s in two independent devices.
In NVIDIA_CUDA_ProgrammingGuide_2.3.pdf, page 75
The use of multiple GPUs as CUDA devices by an application running on a multi-
GPU system is only guaranteed to work if these GPUs are of the same type.
When the system is in SLI mode, all GPUs are accessible via the CUDA driver and
runtime as separate devices, but there are special considerations as described below:
^^^^^^^^^^^
First, an allocation in one CUDA device on one GPU will consume memory on
other GPUs. Because of this, allocations may fail earlier than otherwise expected.
…
It means that GTX295 has 2 GPUs and runs in SLI mode automatically, but 2 GPUs are
regarded as seperate devices, and only 896MB can be allocated for one GPU, right?
So if I want a big memory, then only choice is Tesla C1060 (4GB RAM).
As for Tesla S1070 (with 4 GPUs), even it has 16GB RAM, only 4GB RAM can be used in one GPU.
Is this interpretation right?
Lung Sheng
You have some GTX285 or GTX275 with 2GB RAM, that are worthwile for large datasets, and elsewhere you have to stick with TESLA, but never forget that you wont be able to use the full 4GB and may end-up with only 3GB accessible depending on your usage (and especially if you use Mapped memory).