I have got a bandwidth question.
I got an A10-5800k Trinity APU and I use OpenCL with the integrated GPU Radeon HD 7660D.
Also, I use the Asus F2 A85-M Pro motherboard and 4x 8GB 1866 MHz ram. OS is Windows 7.
I used the AMD BufferBandwidth OpenCL example and ran it on my GPU.
There I got about 25 GB/s for reading. All fine.
Then I changed the BufferBandwidth code to use not nThreads (some calculated number) but maxThreads (maximal threads for the input data). I also changed the input data from 32 MB to 128 MB.
Now I get 70 to 80 GB/s for reading. This sounds somehow too much.
My first theoretical calculations were:
1.866 GT/s (Mem freq) * 256 bit (Radeon Memory Bus) * 2 (Dual Channel) = 111 GB/s
There, 70 to 80 GB/s sound plausible. But normal memory only has 64 bit wide buses. With Dual Channel, that would be only 128 bit meaning only 55 GB/s.
Can anyone clarify the used buses for me, please? Do you really have 80 GB/s on an integrated CPU or is there a bug in the AMD OpenCL BufferBandwidth test? Are there some resources to do some further reading?
Thanks a lot!