What am I supposed to do with your documentation on GPU programming? NeoScrypt has no sole purpose of running perfectly on GPUs either existing or to come. Neither Scrypt had it originally. NeoScrypt performs very well on CPUs now and it's going to be even better with the following optimised releases of CPUminer. It takes full advantage of large and fast L2 caches. This is something you don't find in GPUs and ASICs because high clock and low access latency multiported synchronous SRAM is very expensive. GPUs do very limited caching at all. They rely upon GDDR5 memory bandwidth and large thread count because access latency is terrible. Scrypt ASICs must be using eDRAM which is more expensive than GDDR5 and faster due to lower latency and wider internal data bus, but it's DRAM still and doesn't clock very well. Intel Crystalwell has made it to 1.6GHz which is absolutely not impressive if compared to ~4GHz of modern CPUs with their L1 and L2 caches. Even single ported asynchronous SRAM used for L3 caches seems better than eDRAM for performance reasons. Let's see how GDDR5 and eDRAM work against SRAM under NeoScrypt.

What comes to memory intensity, you can calculate how many quarters it takes to produce a single hash for Scrypt and NeoScrypt.