Our user encountered CUDA_ERROR_OUT_OF_MEMORY on cuCtxCreate call. There are 11 Gb VRAM (and nearly all free), so we think this is not out of memory. What can be the reason?
I find some other mentions about this error on forum, but they encounter it only after ~20 calls to cuCtxCreate.
A little bit more context why we use cuCtxCreate: We were using single GPU for CUDA from two threads independently (with two independent streams) and this approach led to very rare and non-deterministic segfaults in CUDA driver. We suggested that the problem is with non-threadsafe GPU CUDA state, and so we began to use cuCtxCreate to create independent CUDA state for each thread, and this fixed those rare segfaults. But few our users encounter problems on cuCtxCreate call - it fails with CUDA_ERROR_OUT_OF_MEMORY without any visible reasons.
Environment: Windows 10, GTX 1080ti, 384.76.
Bug reported with ID=2023086.
What can be the reason for CUDA_ERROR_OUT_OF_MEMORY on cuCtxCreate? If there is enough memory. :)