at least with a better ratio than the buggy AMD implementation vs scrypt - would have been gentlemanly if they had tipped their hat to Christian for all his effort.
I believe the CPU based validation code has some kind of hardcoded scratchpad buffer size for N=1024. This needs to be fixed.
EDIT: oh look what's there !
Code:
scratch[32768]
EDIT2: okay, a likely fix is now on github. mine away!