No pagable memory access on GTX 1050 Ti under Linux?

Accelerated Computing CUDA CUDA Setup and Installation
1

Hello,

I recently got a GTX 1050 TI in order to try out the new Pascal features. I use CentOS 7 64bit with driver version 375.26 and CUDA 8.

If I query the device properties, concurrentManagedAccess returns 1, however, pageableMemoryAccess just returns 0. Which means, according to the documentation, the pageable memory support introduced with compute capability 6.x is not availabe. Shouldn’t all Pascal GPUs be capable of it? Or is there something to configure?

I compiled a minimal test program which is able to run with cudaMallocManaged() but not with the system allocator malloc(). I compiled it with -gencode arch=compute_61,code=sm_61

Can anybody give me some hints where to look at?

2

If you can oversubscribe the GPU with cudaMallocManaged, then you are getting everything that is expected. Use of the system allocator instead is a feature that requires a specific linux kernel patch (so called HMM patch) which is not conveniently available yet, AFAIK.