AMD Fusion Developer Summit has just started in Seattle, here are some of the initial news, with more to follow
While one can debate the performance pros and cons of the Trinity APU, its tight level of CPU-GPU integration is a fact, including the shared memory controller and such. However, even more is coming, as seen on the keynote presentation two hours ago.
The 2013 Trinity follow-on, Kaveri, will have true shared, unified address space between the CPU and GPU, with common user calls and sharing of data between them. This will eliminate the data copying or wasted bandwith having to reference it indirectly, and come a step forward in integrating the two further. It's supposed to help in both 3-D graphics and the compute.
After the show, a source confirmed that Kaveri will be a true Southern Islands GCN part in the GPU department, without any remnants of the HD6XXX VLIW-4 architecture, and that the Steamroller CPU core parts there would solve a couple of remaining major drawbacks of the Piledriver architecture.
Whether it includes – finally – each core getting its own FPU and instruction scheduler, or something else like common L3 cache for CPU and GPU, non existent till now, was not answered. The point is, Intel's Haswell desktop with supposedly vastly improved graphics (GT3, 40EUs and 64MB cache) – but likely no such high level enhancement like true shared memory – will be out at the same time. So, getting out Kaveri as early as possible should be an imperative for the AMD APU time. And, yes, since the sockets are changed everytime, how about quad memory channel option this time, as DDR4 support is too early then, yet more bandwidth is needed for the GPU at least?