AMD’s John Freuhe posts his perspective on bringing AMD’s Fusion technology — that merging of GPU capabilities with the CPU on a single package. At first blush this seems like an unvarnished good idea, especially for HPC: pulling out the PCI-e bottleneck between the CPU and the GPU will make GPU acceleration much more relevant to a wider variety of applications. But it takes a lot of silicon to do what NVIDIA and ATI are doing today on their chips, and merging CPUs and GPUs will not be a matter of simply gluing the together. There will be tradeoffs in things like capabilities and power budgets between the two, with the potential for different chip personalities for system builders to select from
The only company publicly working on fielding something right now is AMD with their Fusion architecture, and that is only targeted at consumers right now. Everyone in the high end community wants to know when the other shoe will drop. From the tone of Freuhe’s post, it won’t be soon
The need to recompile the application is one reason why you have yet to see heterogeneous computing for servers widely adopted thus far and why I believe the push for an APU within servers is still to come. It’s something AMD is committed to delivering in the years ahead and when the time is right, more info on Fusion APUs for servers will come. For now, AMD is focused on client APUs where GPU acceleration in mainstream applications is happening rapidly, Microsoft Office and Internet Explorer 9 are two of the most recent examples. Optimizing mainstream server applications is not even close to this level yet.
…As you begin to see and hear more buzz about AMD Fusion and AMD Fusion APUs on the client side, just know that there is a plan to bring this level of integration to servers. It’s just a bit more complicated, so we’ll continue to work out the details and expect to see more from us over time.