Bitslice DES S-boxes with LOP3.LUT instructions

[atom]

My first impression is that a < 400 speed boost for 1 high end 3k+ core GPU versus 1 outdated CPU is pretty poor

There's no such general statement. I think your speed, if it includes 3xDES, is almost optimal because 3 x 360 > 1000. DES is a special case because of the sbox lookups. Not the memory access is slow, GPU memory is extreme fast, but for each request the pointer calculation creates three additional instructions to pinpoint the correct section in memory. Other hashes like MD5 which do not require such lookups are much faster.