CUDA reports no devices present when logging in over SSH

Question

I'm on Windows 7, CUDA 5, but I also experienced this with CUDA 4.2 (prompting an upgrade to 5.)

I remote in over SSH (using cygwin's ssh), and when I run the following test program, it tells me no device present.

$ ./a.exe
CUDA error: no CUDA-capable device is detected

$ cat test.cu
#include <stdio.h>

__global__ void foo()
{
}

int main()
{
  foo<<<1,1>>>();
  printf("CUDA error: %s\n", cudaGetErrorString(cudaGetLastError()));
  return 0;
}

When I try to run deviceQuery, it loops infinitely. Here's a snippet:

CUDA Device #91366
Major revision number:         0
Minor revision number:         0
Name:                          pqrstuvwB
Total global memory:           1895825521
Total shared memory per block: 3
Total registers per block:     -1073741632
Warp size:                     0
Maximum memory pitch:          132
Maximum threads per block:     -1819111168
Maximum dimension 0 of block:  -1751738988
Maximum dimension 1 of block:  1997970554
Maximum dimension 2 of block:  0
Maximum dimension 0 of grid:   1895825521
Maximum dimension 1 of grid:   0
Maximum dimension 2 of grid:   0
Clock rate:                    0
Total constant memory:         0
Texture alignment:             0
Concurrent copy and execution: Yes
Number of multiprocessors:     0
Kernel execution timeout:      No

These programs give the expected output when run locally. I have two GTX 480's and a 9500 GT. In a prior version (I guess pre-4.2) this did work, and I used it for developing CUDA apps remotely. Now I cannot.

I saw some remarks about CUDA not working over Remote Desktop, and suggesting people switch to VNC - but nothing about SSH. For what it's worth, it happens when the monitors are both on and off.