Right, the lcache is used in scrypt. Reading to and writing from the RAM would be too time intense. So forcing the process to use enough memory does it make harder for GPUs,FPGAs,ASICs to outperfom the lower processors. Anyway: I think you guys are missing the point here. For getting a pure proof of work mechanism done there is no need for an enc/dec function like scrypt. It can be done with simple hashing (what we are looking for). Of course it's interesting to see how scrypt makes the processing cache intense and it may be something we should include in our hashing function too. The GPU resistance should not only depend on simply high memory usage. We should have a hashing function that includes several different ways to make it CPU friendly. The question is still the same: Anything else than radix sort we should include?

Related to the botnet and co resistance I've already mentioned that this is a topic for later. Having a good enough hashing function has to be done first. It's about getting this done step by step. Don't expect another useless altcoin that just clones an existing one with little changes in it.