“Kaveri changes everything”, said Nic. “The level of flexibility and cooperation between the CPU and GPU areas within a system are now on a different level”.
He explained, “With the new APUs, we can offer up to 4 unlocked CPU cores, each of which have been designed to boost computational efficiency, feed the cores faster and improve single core execution – while at the same time providing up to 8 supercharged GPU cores, based on the latest GCN architecture, for maximum graphics performance”.
“AMD's Heterogeneous System Architecture offers amongst other things, accelerated media playback with UltraHD video encode/decode built into the GPU areas, which also allows up to improve video editing, while at the same time having enough power to offer next generation NUIs (Natural User Interfaces)”, Nic told us.
We asked if the improvements are localised at one price point, for example in the top of the range chips.
“Far from it”, said Nic. “While performance in PC Mark 8 v2 is 8% better when comparing the best Kaveri chips to the best Richland chips from the last generation, the improvement in the sub-45w area is much bigger – as much as 15% faster”.
That's in raw performance, but what about ‘Performance Per Watt'?
“When you measure system scores in 3DMark Fire Strike, you will see over 33% better performance per watt in the mid and high-end chips, but the low end increases by as much as 75%”, he told us.
Intel can start HSA using their own gpu chips which are poor but still something. They can surely create a much better memory controller and implement a scaled down eDRAM writeback cache to speed up DDR3 accesses much like the Xbox 1 did. In fact, Iris PRO already has it implemented but with 128MB which is too big and eats 20watts of power constantly. Using 32MB eDRAM is enough to make a serious bite to HSA performance they will aim at.