AMD just released its reference guide (PDF) for generation three of its Graphics Core Next (GCN) architecture, also known as Volcanic Islands. The Reading through the document offers some insights into exactly how GCN3 works.
The document also resolves a bit of a mystery about AMD's Carrizo APU—namely, what generation of AMD's graphics technology is built into this mobile processor, which is due to launch mid-year.
We know Carrizo implements context switching for compute kernels, a feature required for compliance with AMD's HSA 1.0 spec for converged CPU-and-GPU computing. Right up front, in a section titled "Important differences between Generation 2 and 3 GPUs," the reference guide mentions this feature: "Compute kernel context switching." Search the document for "Carrizo," and you'll find 24 instances. Many of those references have to do with XNACK_MASK_LO and XNACK_MASK_HI, special instructions that are marked "Carrizo APU only."
At the very least, Carrizo looks to be using Volcanic Islands-era shader cores, and those cores include specific provisions for converged computing. We don't yet know whether Carrizo's IGP includes the improved frame-buffer compression first outed in the Tonga GPU that drives the Radeon R9 285. Compression of this sort could make Carrizo's graphics unusually effective within the constraints of a CPU's memory subsystem.
The docs even provide some logical block diagrams of Volcanic Islands-series GPU. Behold:
At first blush, I'm not sure we're seeing anything in particular in these images that suggests monumental differences between GCN3 and prior generations of AMD's GCN architecture. Perhaps you all will spot something I haven't yet.