Originally Posted by SCollins
Originally Posted by KyadCK
Not really since the PCI-e bus latency to the GPU is the number one killer of GPGPU performance.
A stock Llano 3870k matches a 8350 at 4.8 + 7970 in espresso encode on personal experiance. hUMA and hQ in Kaveri will reduce that latency even more, allowing for seriously improved GPGPU processing.
It's kinda why they're going this route in the first place...
some of that pci latency is os based to, I mean, we are talking nano seconds here, not microseconds, seconds etc.
And Application side (which HSA addresses), and hardware communication side (which hQ addresses), and memory copy time (which hUMA addresses)...
Really the entirety of HSA is a massive boost for GPGPU workloads. Espresso never really got to use my 7970 more then 12-16% because it was too busy getting work to and from the GPU to use much more of it. It was still faster to use the GPU too, but the advantage wasn't very large and the APU was able to keep up.