But you can still gain linear increases in matches with just increasing the hashrate so at the very least you can just have an ASIC or GPU that hashes and sends the results to a CPU that manages the memory and checks for matches. GPUs will still be crazy good for it since they could hash 100s of times faster than the CPU and with 2-3GB of memory on board, the CPU would need too much memory to pull even. I think a GPU/CPU hybrid would be best in class until ASICs with the GPU doing mostly hashing and then sending the results to system RAM and the CPU which checks for matches.
With this linear increase just by increasing the hashrate the Momentum ASICs will be at least as good as the scrypt and probably much better since they could just extend an off the self system with lots of RAM and use a CPU to manage the hashtable and match checking part. Design might be a bit harder, but that's just a short term concern.
The reason it is GPU resistant has to do with bus speeds and the unpredictable runtime of my scrypt-like hash function. I did not discuss this in the white paper which was meant to be more generic. I also use hardware accelerated AES to give the CPU an edge over GPU.