[BrynS]

ExtremeTech has posted a podcast interview (21.3MB) with Phil Hester (AMD's Chief Technology Officer) covering Barcelona, Fusion, CPU core scaling, etc -- the interview starts at 14m:50s and runs for ~30 minutes.

It's fairly general and not too detailed, but there were some bits and basic affirmations on Fusion, power/thermal reductions and the prevalence of standard/custom logic in future GPUs.

[Rys]

I had a listen and Hester talks about a time where the CPU and GPU are blurred in terms of the instruction set and programming model. He initially starts to talk about Fusion, but segues into talking about a system where you might run part of a 'CPU' thread on a GPU in the system, and vice versa, depending on load, a dispatch subsystem and what have you.

He specifically references x86 there, too, if I'm not mistaken, hinting at specific x86 execution support on the graphics processor(s) at some point. While it's just a vision, given he's the CTO you can probably assume that AMD are looking at that going forward.

[3dilettante]

At the rate AMD's going with Fusion, we're probably looking at 2010-2011 for the integration to reach that point.

It seems doable, once the GPU and CPU are close enough and no longer separated by a peripheral bus.

The command processor already does some work dealing with contexts and virtualization, so it must have some resources of its own.
I'm curious about what actually is contained by that unit.

It seems plausible that the equivalent of a stripped-down 486 could be integrated with the command processor. That would provide a jumping-off point for compatibility, especially when dealing with code where exceptions and interrupts are to be handled correctly.

It would be wasteful to apply full compatibility to the rest of the GPU when the command processor could synthesize and dispatch the necessary instructions for code with less stringent requirements.


edit:
I kind of wish I had speakers on my current machine so I could hear if there were any answers or partial answers to my questions.