AMD Pushes HSA For Data Center Servers

Chipmaker AMD is pushing a new architecture that allows CPUs and graphical processing units (GPUs) to work together to improve the performance of servers and other devices.

The idea behind what AMD calls its Heterogeneous System Architecture (HSA) is that when the central processing unit (CPU) and graphics processing unit (GPU) operate as one, they work better together for faster processing at lower power consumption than they would as separate processing elements. HSA eliminates the need for an application to copy data back and forth between the CPU and GPU, enabling applications to run faster, according to AMD.

The framework is designed to take advantage of AMD's Accelerated Processing Unit (APU), which is a processor that brings a CPU and GPU together onto one die.

AMD says the HSA platform works across device types, including servers, smartphones, tablets and gaming consoles. At its developers conference this week in San Jose, Calif., the company announced that its new “Berlin” line of APUs, which is built on HSA, will be available early next year.

APUs based on HSA will meet growing data center demand for faster servers that also use less power, says John Williams, VP of server marketing and business development at AMD.

“The fundamental landscape is changing. There are a couple of things that have to happen,” said Williams in an interview. “One is that we have to rethink the architecture of the data center itself. We have to start looking at high-density servers, [and] we have to look at more power-efficient solutions."

The HSA platform should appeal to data center operators who are having to use more GPUs, not just for graphics but for all sorts of high-volume workloads such as data reduction and signal processing, says Richard Fichera, a VP and principal analyst at Forrester Research.

“I’ve talked to clients who’ve got several thousand nodes of multiple GPU servers in a data center. Those consume a huge amount of power, so the HSA architecture ... has potentially huge ramifications for specialized workloads,” says Fichera.

Fichera adds that Intel has Xeon 5 chips that have some HSA-type capability but lacks the level of integration of CPU and GPU that AMD is now presenting.

[Read about Intel's plans for Atom SoCs designed to penetrate all areas of data center infrastructure in "Intel Unveils Plans To Dominate The Data Center."]

Since this was a developer conference, AMD introduced a number of tools to make it easier for developers to write code based on HSA. It introduced an SDK, APP SDK 2.9, and an upgrade of its CodeXL developer tools (version 1.3) for Windows and Linux systems.

Another design quality of HSA is that apps can be written to operate across device platforms without having to be rewritten to run on smartphones, desktops, gaming consoles or servers, notes Forrester’s Fichera.

“HSA has been designed from the very beginning to be instruction-set-agnostic. So it doesn’t depend on it being an x86 processor or an Intel processor or an AMD processor, and that’s a really significant leap,” Fichera explains. “Up until now, all of these schemes to share resources have all been very specific to certain vendors or technologies. A potentially a huge feature of HSA is that it is instruction-set-agnostic.”