Large memory allocation with CudaHostAlloc fails with CUDA 8.0 release build

True that. I mixed up the date format (assuming 10/06/2016 to be June), sorry. I have corrected my posts. Thanks for clarifying this!

Latest driver 377.35 still does not fix it. Does anyone know how to get Release Notes for a specific driver/OS? I googled but could not find Release Notes for 377.35/Tesla/Windows Server 2012 R2.

As of last week’s CUDA v9.0 release (9.0.176) this bug seems to have been fixed. (In fact, the bug seems to have disappeared in the pre-release versions of CUDA 9 as well.)

Almost a year has passed since this bug was first reported, but NVidia has never acknowledged that the bug existed or updated the bug-report status (https://developer.nvidia.com/nvidia_bug/1824366), so one cannot be certain whether the current CUDA release contains an actual bug fix or just a lucky rearrangement of code and data. We’ll see…

It is not fixed here. I use 9.1.84 on a system with 2 GTX 1060 3gb. I have 500meg left on device when I try to allocate 2.5G on host using cudaMallocHost for buffer transfer. Very disappointed…