AMD: Era of heterogenous multicore processors on the horizon

Yesterday I met with AMD and Microsoft to see a demo of Virtual Server running on an AMD-based PC. The software, Microsoft's Virtual Server R2 running an as-yet unreleased service pack, takes advantage of AMD's chip-level support for virtualization, allowing Virtual Server to run more efficiently.

But it was my side discussion with AMD development manager Andrew Shajenko that piqued my interest. In the future, he says, computer processors will contain a heterogenous mix of task-optimized cores. These will be logically aggregated into resource pools and assigned to service specific tasks. For example, a video streaming application might go to a core that specializes in floating point computation, while a voice over IP application might run on a core optimized for digital signal processing.

These specialized cores sound a bit like co-processors, Shajenko acknowledged, but the approach is very different. They will be separate cores within the processor chip or chip set. And they could be aggregated by function across servers or groups of servers to form virtualized resource pools (although operating systems would have to be modified to support that). You can think of it as virtualizaton at a more granular, lower level, although "It's really resource mangement," he says.

I asked Shajenko whether we would see servers with homogenous groups of task-specific cores that could then be easily aggregated into, say, groups of task-specific blades, or whether each server would contain processors with a heterogenous mix of cores. It could go either way, he says. "It really depends on [for] what group of applications a server is intended. If a server is not using certain capabilities of a chip there's no reason to have a chip with those capabilities," he says.

"The biggest issue is if you provide a certain service on a core, then for all applications you have to save and restore contexts. This is a way to cut out that step of having to save context from one application to another." Processors, he says, will get better at offloading software tasks into hardware.

While Shajenko says this move to task-specific cores is the general trend, any product plans in this direction are still a few years away. " This is not part of a roadmap. It's a hint of where things may be going," he says.