At least 8 KB of memory are needed for groestl. Not using global ram doesn't equal register only. In my case I'm using part local ram and part memory cache. Same for hetpas.
With bitslice you can reduce memory usage further but not completely.
Still asic can be done if enough volume, which is not the case with DMD IMHO. And with the other coins of the same algo which are dead or almost dead it's even more so.