The upcoming Fusion processors raises some questions from the OpenCL point of view. Maybe it's too early to start worrying about it but perhaps this is the chance developers have to express expectations in an earlier stage.
1 - Will the GPU part be able to have direct access to RAM memory? Currently it is known that data transfer to OpenCL Device is very time-consuming.
2 - Will OpenCL see Fusion's GPU and CPU as two independent devices? I would think so because it may be hard to tell which code works best on CPU and GPU.
3 - Will the GPU and CPU share L1/L2 cache? Again, we may want to manipulate buffers using the GPU and then recover them directly in a serial code which is running on the CPU.
4 - If a Fusion processor has, say, 4 CPU cores and 5 GPU workgroups, will it be possible to use 1 core as scheduler and use device fission to use each remaining core/workgroup separatedly? I imagine that in the future multiple applications will use the GPU and CPU and none of them should consume all resources.