AMD sows the seeds for hybrid CPU/GPU systems before January launch

  • Jon Gold (Network World)
  • 11 November, 2013 20:05

AMD's new "Berlin" server APU will bring an innovative technology called HSA, or heterogeneous system architecture, when it launches in early 2014, and the company is already working hard to lay groundwork for its adoption.

The essential idea of HSA is to create a layer of abstraction that treats the CPU and GPU as a single processing entity allowing the GPU's logical muscle to be applied to tasks where it's useful, and eliminating the need for the two processors to write data back and forth to each other. This means that a computer can theoretically operate far more efficiently with HSA than without, getting much more performance out of any given hardware setup.

[MORE SERVERS:Cisco puts its unified computing servers into OpenStack clouds]

AMD says that it's seen excitingly outsized performance gains on some metrics including a face detection algorithm used on digital video that operated at more than twice the speed, using less than half the power, on an HSA-enabled system when compared to identical hardware structured traditionally.

Senior Product Marketing Manager Guy Ludden says that the main focus remains maximizing bang for the buck.

"You're really going for price/performance," he says. "It's not pure computing power with these initial APUs."

The undeniably promising technology has a number of obstacles to surmount before breaking into the mainstream, however, not least of which will be convincing software developers to get on board. While HSA should make development simpler in the long run, it still represents a short-term change.

This is why AMD will be playing up both new and existing efforts to lay the software foundations for HSA at its developer conference this week. Ludden gave several examples of this groundwork that have already been put in place, including projects like APARAPI. That particular project apes the general idea of HSA by converting Java bytecode into OpenCL and running it on a GPU.

APARAPI will be written directly into the OpenJDK by version 8 in 2015, according to Ludden.

"Just think of what that opens it up to," he says. "That's a huge step in the right direction."

Ludden also stressed AMD's commitment to open standards, highlighting the company's central role in the foundation of the HSA Foundation, and talking up its contributions to open-source software.

"Most of what we're working on is actually going to be open-source," he says. "We've always tried to work with the community, and that's why we're doing it with the HSA Foundation."

"Berlin" will blend AMD's new Steamroller CPU cores with existing Sea Islands GPU cores, allowing both to work together under the coordination of the HSA layer. An ARM version, codenamed "Seattle" is scheduled for later this year.

Email Jon Gold at and follow him on Twitter at @NWWJonGold.

Read more about data center in Network World's Data Center section.