I tried it on Linux, and it fails to malloc on this line

  error = cudaMalloc((void **)&dev_countbits, sizeof(uint32_t)*NUM_COUNTBITS_WORDS);

My card is a compute 2.1 GTX 460 with only 768MB ram and currently running two screens. I have changed arch to sm_21

FYI "Device has 647233536 free of 804454400 total bytes of memory"

Suggestions?