I just have a question about an issue that is currently coming up. We have waited a very long time so that we can switch to VS2015 because only CUDA 8 does support the compiler.
But now another issue comes up, because the GPUs we have our at our customer sites are mostly C2050 and C2075, but also many other cards and I can only download drivers for CUDA 7.5 for both cards.
Our framework is following: We are working in a C# #environment where we build PTX code from our CUDA code. We currently build for compute capability 2.0 so that we can support most of the cards out there. The PTX code is then JITted at runtime and we access all CUDA methods using the driver API.
Since we are using VS2015 and have to install CUDA 8 the question is, if our new versions will still run on these systems if I have to update the drivers in the field? Or do I have to update them at all since we are running on the almost lowest capabilities the GPU can handle (that would be the best option for us).
In general I think there are a lot of complaints where the CUDA toolkit version is so tightly bound to a Visual Studio version, so either CUDA should get rid of this restriction or - if someone can suggest another compiler enviroment I would be glad to install it without having to live with this restriction.
If you use CUDA 8, any system you run on must have a CUDA 8 compatible driver.
A CUDA 8 compatible driver is contained in the CUDA 8 toolkit installer.
Understood. But this means C2050 and C2075 cannot be used anymore when I switch to VS2015?
I don’t think that is correct.
The driver bundled in the CUDA 8 toolkit installer should suport C20xx, regardless of which supported version of VS you use.
thanks for your response:
Maybe you are right. But at least it seems to me that there might be a problem since I have to choose the CUDA toolkit version when downloading the driver.
For a C2050 the driver is only supported until 7.5, I cannot download CUDA 8 toolkit.
See here: http://www.nvidia.com/Download/index.aspx
Nevertheless I’m also not quite sure why the toolkit is attached to the driver.
I made no comments about downloading the driver.
All of my comments referred to the driver included in the CUDA 8 toolkit installer.
It may well be that separately downloadable drivers do not exist (yet) in the driver branches for CUDA 8, which list C20xx devices as supported. My guess would be that those are coming, but I can’t comment directly on the future.
Nevertheless, to repeat, the driver bundled in the CUDA 8 toolkit installer should support C20xx devices, indicating that the general concept is possible, if you use the “correct” driver.
Driver installers are generally extractable from the toolkit installer, although I haven’t done this lately using the windows toolkit installer, so at least in theory it should be possible to use that driver installer to update whatever c20xx machines need to be updated, in order to run CUDA 8-built codes on them.
If you want a separately downloadable driver, you may have to wait. I haven’t looked or scoured the website ATM.
The CUDA 8 windows toolkit installer can also be used as a driver-only installer by simply selecting the “advanced” options and selecting the components you wish to install (e.g. driver components only).