I wanted to try out CUDA 6.5, but the installer warns me that I’ll need a driver from the 340 series to do so. However for Tesla cards the highest available driver version is 331.89 ( http://www.nvidia.com/download/driverResults.aspx/77204/en-us ).
I installed CUDA 6.5 anyway to see if I can get it to work, but my programs fail with “undefined symbol: cuMemHostRegister_v2”.
So I’m guessing I do indeed need a 340 driver. Is there an ETA for such a driver, or does anyone know whether the current 340.xx drivers do unofficially support Tesla cards?
The previous CUDA 6.5RC runfile installer for linux, or unified installer for windows, included an r340 driver. I’m sure the CUDA 6.5 production release runfile installer includes the necessary driver as well.
So for now I would use the unified installers.
You’re right, the installer did contain 340.29, but I wasn’t sure whether that would work on Tesla cards. Turns out that it does. Which I think is pretty weird, considering that you can’t find a 340.xx driver yet as a single download (see link posted above).
I confirm, this problem still exists. My cluster has some nodes with Intel Xeon CPUs and NVidia Tesla K20m boards.
NVidia driver 340.29 does not contain CUDA-capable driver. Applicarions that were compiled with CUDA 6.0 crash with segfault. However, they crash if the old driver (331.89) is installed too.
NVidia tech guys, could you help us? How soon will a new CUDA-capable driver be released?
Applications that were compiled with CUDA 6.0 may need to be recompiled under CUDA 6.5. This is not a driver issue. The exception would be applications that were compiled with static linking to CUDART and any other CUDA libraries that depend on CUDART. Unless these steps were taken, the dynamically loaded 6.5 CUDART is not compatible with applications that were compiled (linked) against the 6.0 CUDART.