SAN JOSE, Calif.—Much of Advanced Micro Devices' focus the past couple of years on its accelerated processing unit strategy focused on desktop PCs and notebooks and, more recently, embedded systems.
However, AMD officials expect to bring the APU architecture—where the CPU and graphics are integrated onto the same silicon—into the server space in 2014, starting with the release of the x86 Opteron X-Series "Berlin" APUs in the first half of the year. At the AMD Developer Summit 2013 here, the officials are laying the software groundwork in preparation for Berlin, a family of low-power chips aimed at cloud computing and hyperscale data center environments, where energy-efficiency is needed.
The recurring theme of AMD's show, which runs through Nov. 13, is the company's heterogeneous computing push—the ability to create computing platforms that leverage both the CPU and the GPU, depending on the workload, and the ability to easily move those workloads between the two. Heterogeneous computing answers the growing demand for more compute performance and more power efficiency in servers, demands that are driven by such trends as cloud computing, big data, mobility, video and virtualization.
AMD and its partners in the Heterogeneous System Architecture (HSA) Foundation also want to make it easier for developers to write programs that take advantage of the capabilities in these new systems—not only the ability to handle the growing number of parallel computing workloads, but also to easily move workloads between the CPU and GPU.
"The key to accelerating the workloads [is they] need to flow freely between the CPU and GPU," Phil Rogers, AMD Corporate Fellow and president of the HSA Foundation, said during a keynote here Nov. 11 while outlining the software efforts the company is making around heterogeneous computing.
The Berlin APUs will be built on HSA Foundation specifications, which includes the ability to view the GPU and the CPU as a single processor, and to move the workloads to whichever one is most needed. In addition, programmers writing code will no longer need to be concerned with determining whether the software will run on the CPU or GPU, Rogers said. They can simply create the software and the system will make the decision.
The HSA specs also give GPUs more access to memory, according to AMD officials.
In addition, AMD also is adding CodeXL 1.3 to server APUs. CodeXL 1.3 is the vendor's developer tools suite for Windows and Linux that includes remote debugging and profiling. Rogers also unveiled Project Sumatra, an open-source project with Oracle that enables Java developers to leverage the GPU capabilities in the APUs. In a keynote here Nov. 12, Nandini Ramani, vice president for the Java Platform at Oracle, said Java—given the size of the developer community, the reach of the programming language and Oracle's efforts to embrace parallel processing—is a good fit for AMD's heterogeneous computing efforts. There are 9 million Java developers worldwide, 80 percent of mobile developers use Java and more than 3 billion mobile devices run Java, Ramani said.
In addition, Rogers also talked about AMD's work with SUSE on the GCC/HSA Project, which will bring HSA support to the GCC open-source Linux compiler, while AMD's OpenCL math libraries—which were given to the open-source community in August under the clMath moniker, will enable programmers to more easily run their high-end scientific computations on both APUs and discrete GPUs from AMD.
AMD officials see the dense server space as a significant growth opportunity for the company. Along with the upcoming Berlin APUs, AMD next year will begin building server chips based on ARM's upcoming 64-bit architecture, enabling the vendor to offer system makers a choice of platforms. At the show, Hewlett-Packard officials announced the company next year will start selling AMD-powered versions of its energy-efficient Moonshot servers.
AMD this year launched its x86 "Kyoto" Opteron processors for microservers, and also is addressing the market through its SeaMicro server business.
ARM and its partners—including Calxeda, Applied Micro and Marvell Technologies—see such microservers as an opportunity to bring ARM's low-power system-on-a-chip (SoC) architecture into the data center. ARM-based 32-bit chips can be found in most smartphones and tablets, as well as a large number of embedded systems. Now the chip designer wants to challenge Intel in the low-power server market.
Intel already offers low-power Atom SoCs for microservers. The first of HP's Moonshot systems launched with Intel inside.