I am eager to try multi-GPU computing. For a long time, I was using GTX TITAN, now I have a GTX 1660 as well, latter as a primary GPU. Everything works, if I have one of these GPUs installed. If I have both, I can see both in tools like GPU-Z, but only the 1660 is listed anywhere as capable of CUDA or OpenCL functions. I’m developing on Windows using mingw and NVIDIA CUDA toolkit.
Why could this be? And if I get a different secondary GPU (say, another 1660), is there any guarantee of more success?
GTX Titan is a compute capability 3.5 GPU. The most recent CUDA versions and GPU driver versions no longer support that GPU. If you load a driver on your system that is 470.xx or older (don’t go a lot older), you should be able to see and use both GPUs.
mingw is not a supported development environment for CUDA on windows. The expectation is that you would use Visual Studio. I’m sure its possible to use basic mingw tools like editors, etc, but attempting to use the mingw compiler as a host compiler for CUDA is not supported and not recommended.
And, if it were me, I would not attempt to use any CUDA version newer than 11.4 for that combo of GPUs.
Thank you! That explains why the driver version number jumped by quite a bit when I installed the 1660 and the newest driver for it. Seems that unfortunately these cards are lumped under one driver.
Yes, for both windows and linux, all serviceable NVIDIA GPUs in a single system/OS are (must be) serviced by a single driver. Therefore, supporting an older (cc3.5) and a newer (cc7.5) GPU in the same system often requires finding this kind of “driver sweet spot”. In some cases its not possible. For example, supporting a cc2.x GPU (Fermi generation) and a cc7.5 GPU in the same system is not possible. There is no (single) driver released by NVIDIA that supports that range of compute capabilities. And your cc3.5 GPU is already deprecated and removed (as I already indicated) in the newest driver and CUDA versions.