Quote:
Originally Posted by erunion 
No, its not.
Using OpenCL(Why didn't wiki mention this feature? Is it actually implemented yet?) is not enough to be HSA. APUs and Ivy Bridge can already run OpenCL on their iGPUs. But that is just the CPU telling the GPU what to do. That has massive performance overhead. Its only useful on a large, very parallel task. Latency doesn't matter on a file compression that takes several minutes.
Integrating the GPU as a coprocessor on the hardware level removes those performance overheads.
Apple uses Intel chips so will get those benefits when everyone else does. So again, that has nothing to do with this.

No, its not.
Using OpenCL(Why didn't wiki mention this feature? Is it actually implemented yet?) is not enough to be HSA. APUs and Ivy Bridge can already run OpenCL on their iGPUs. But that is just the CPU telling the GPU what to do. That has massive performance overhead. Its only useful on a large, very parallel task. Latency doesn't matter on a file compression that takes several minutes.
Integrating the GPU as a coprocessor on the hardware level removes those performance overheads.
Apple uses Intel chips so will get those benefits when everyone else does. So again, that has nothing to do with this.
I'm not trying to argue about it.
The point of GCD is to make it easier to program for multiple cores AND OpenCL aka GPGPU.
So for example: I want to write a simple app that uses a lot of math, some of which is highly paralleled. I can program my app to use GCD instead of programming the threads directly. Then GCD will query the system for what hardware it has and use the appropriate hardware for the task.
In the case of the highly paralleled math, it might query for whether or not a discrete OpenCL device, aka a graphics card, is installed. Then it will select whether it should use the GPU, or the CPU to perform the task at hand.
See OpenCL can be done on a CPU, but is better on a GPU, so it looks for the most appropriate one and uses it.
That sounds pretty much the same to me as what the AMD marketing guy was selling. And since they are courting software companies to develop it, it won't be done in hardware. At least not fully.
.02












