It easy to understand the hUMA.
Today when you have a task in PC that need to use both GPU and CPU you write in a simple term a code like:
Read GPU Memory
GPU Work
Write GPU Memory
Copy from GPU Memory to CPU Memory
Read CPU Memory
CPU Work
Write CPU Memory
Copy from CPU Memory to GPU Memory
Even if a Unified Memory Pool you need to copy the data between the CPU allocated space and GPU allocated space
Read Memory (GPU allocated)
GPU Work
Write Memory (GPU allocated)
Copy from GPU allocated memory to CPU allocated memory
Read Memory (CPU allocated)
CPU Work
Write Memory (CPU allocated)
Copy from CPU allocated memory to GPU allocated memory
With hUMA you kill these copy tasks from CPU/GPU.
Read Memory
GPU Work
Write Memory
Read Memory
CPU Work
Write Memory
Due lattencies issues with the memory the gain removing these Copy between memory part five a significative boost in performance of the code created to run the game and open more possibilities to developers uses more taks that combines GPU and CPU.
A good exemplo... sometime a part of a task runs better on CPU than GPU but the data used are in GPU memory so the dev choose to use the GPU to do this task because the gain in perforance won't be significative over the copy/move data between the memories but now with hUMA will can use the CPU and GPU without need to move/copy the data... so it a task run better on CPU or GPU they use it without need extra work.
That's what I get from hUMA.
PS. My example is too simple and not like it is implementend... just to make a point.
The PS4 GPU is created with tech that AMD will release on the market with GCN 2.0 next year... so it is safe to say that for a time the PS4 GPU will be the most advanced GPU in the market... even PC can't have that yet.
