nvcc die in -dlink stage to create a static library


I’m using CentOS 7 and cuda 6.5 with mvapich2.
I’m creating a static library but in last days when creating the dlink.o
the nvcc dies with the error:
nvcc error : ‘nvlink’ died due to signal 11 (Invalid memory reference)

Has anyone experienced the same behavior?

In last tests I made, it seems to be related with the combined size of the .o files,
If the size of all files is bigger than a certain value, the nvcc dies…
Up to 43 MB there if no problems, but at 46 MB the nvcc dies…

Currently the cuda 7.0 is not installed on this machine and therefore I cannot test it.

I have successfully linked in a Ubuntu machine, 14.10, with the same mvapich2 version as well as for the cuda 6.5 and 7.0.


Probably best to create a short code that reproduces the problem, then file a bug at developer.nvidia.com