I've just run the PCIeSpeedTest v0.2 tool that can be found on AMDs website (somewhere) and just concerned about the low GPU-CPU transfer speeds on decently large buffers..
I'm PCIe.2.0x16 but I can get ~4.9Gbps CPU -> GPU, which is fine. However only on 131072byte transfer size can I reach 4Gbps on GPU->CPU. Anything from 262144 bytes and above the transfer drops to ~500Mbps GPU->CPU, even transferring large buffers several hundread MBs. Just to re-iterate from 16bytes all the way to 131072bytes the GPU->CPU transfer rate increases linearly with buffer size (as expected), topping out at 4.54Gps, so there's nothing to suggest it's physically (hardware) limited in any way.
Just curious is this a driver/hardware issue, or just a fault of the tool..?.I've not tried any actual real life tests of my own in OpenCL. If I recall correctly however, even 47xx devices can manage almost 1Gbps GPU->CPU using this tool.
I'm running Windows 7 x64, Catalst 12.2, SDK 2.6, Intel i7 870, ASROCK Extreme P55 motherboard.
Have you tried running BufferBandwidth from the SDK? It's much more recent and, hopefully, more reliable.