I've written a set of OpenCL kernels. When testing on an AMD R9 Nano with 4 GB memory, the kernels execute correctly. When testing on a Nvidia GTX 1080 with 8 GB of memory, the kernels fail to execute unless I reduce the buffer memory by roughly 8 times. Then, they operate correctly. There are several buffers, but the largest for the working Nvidia example is ~45 MB.
CL_DEVICE_MAX_MEM_ALLOC_SIZE is 3 GB for the AMD card and 2 GB for the Nvidia card, but this seems too small a difference to cause the problem.
Previously, there were known problems with Nvidia CL_DEVICE_ADDRESS_BITS being 32 bit rather than 64 bit, but that appears to no longer be correct.
Are there any current known differences in how Nvidia and AMD handle memory buffers in OpenCL that might cause this problem?