I have some legacy software that we compile in CUDA 9.2 then is deployed in CUDA 9.2 on Tesla k80s (compute arch 3.0).
Our new architecture utilizes Tesla V100s (compute arch 7.0). We still compile in 9.2, but deploy on 11.0.
We compile using architecture specific (300 700) .cubin
We are getting a CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES = 701 error on one of our kernel calls. The code and arguments are unchanged…
Could it be related to the version mismatch?