IBM, Xilinx to Expand FPGA Use in Power, OpenPower Systems

The partnership, announced at the SC 15 supercomputing show, is aimed at bringing acceleration technologies to mainstream data center workloads.

IBM logo

IBM is partnering with Xilinx to expand the use of programmable chips to its Power systems as the tech vendor looks to gain more share of the server chip market from Intel.

The two companies will collaborate on a wide range of efforts that are focused on using Xilinx field-programmable gate arrays (FPGAs) in Power systems to help speed up such workloads as big data analytics, machine learning, network-functions virtualization (NFV), high-performance computing (HPC) and genomics. The partnership will touch on everything from infrastructure to middleware to software.

Accelerators such as GPUs and FPGAs have been used in the HPC space for almost a decade as a way of increasing the performance of the systems while holding down power consumption. However, with new workloads that are demanding faster processing and quicker results—such as data analytics, where organizations are looking for faster and better insights into the massive amounts of data being generated—the demand for acceleration technologies is expanding beyond HPC and supercomputers, according to Brad McCredie, IBM Fellow and president of the OpenPower Foundation.

"Now it's going to migrate to mainstream workloads," McCredie told eWEEK. "We are seeing it move into the enterprise. It's a clear trend."

The bulk of the focus on accelerators over the past 10 years has been on GPUs from Nvidia and Advanced Micro Devices and x86-based Xeon Phi coprocessors from Intel. For example, the number of systems on the twice-yearly Top500 list of the world's fastest supercomputers that use such accelerators has been growing. In the most recent list released Nov. 16 at the SC 15 supercomputing show in Austin, Texas, more than 20 percent—or 104—of the systems use the GPUs or Xeon Phi accelerators. IBM has been using Nvidia GPUs as accelerators for several years.

However, the use of accelerators—and the kinds of accelerators—is expanding, with FPGAs in the spotlight. Intel is buying FPGA maker Altera for $16.7 billion to help it expand in such areas as cloud computing and the Internet of things. Qualcomm has begun sampling ARM-based server chips that use Xilinx FPGAs. FPGAs can be programmed through software and are becoming increasingly important accelerators for cloud and Web-scale environment.

"You need lots of partners and lots of kinds of accelerators to meet the workload needs out there," McCredie said.

The multiyear collaboration with Xilinx was one of several acceleration-focused announcements coming out Nov. 16 from IBM and OpenPower at the SC 15 show. For example, IBM also announced that it using Nvidia's Tesla Accelerated Computing Platform in its Watson platform. In addition, E4 Computer Engineering and Penguin Computing—both OpenPower members—are releasing new systems based on the OpenPower design concept, which includes running Power8 chips and Tesla GPUs.

IBM's partnership with Xilinx—which was developed through the OpenPower Foundation—hits on a number of areas. IBM Systems Group developers will build solution stacks for Power-based servers, storage and middleware systems that will leverage Xilinx FPGAs for such workloads as OpenStack, Docker and Spark. Big Blue also will develop Xilinx accelerator boards that can be used in its Power systems, while Xilinx will offer a Power-based version of its software-defined SDAccel Development Environment and libraries for OpenPower developers.

In addition, the companies will develop products that use IBM's Coherent Accelerator Processor Interface (CAPI) technology, which further helps performance by enabling components to create and directly access large memory spaces on the Power processors.

Both IBM and ARM—as well as ARM's myriad chip making partners—are looking to capture some of the 95 percent market share that Intel has in the server chip space. Both are taking the approach that a larger partner ecosystem can help drive innovation faster than Intel's relatively closed chip development and manufacturing process. IBM for decades used Power chips on its own systems, but in 2013 launched OpenPower with the idea of extending the reach of the Power architecture into other market segments.

The foundation now has more than 160 members, and Xilinx increased its membership position within the group.

Analysts have said that end users are hoping for increased competition for Intel for a range of reasons, including the belief that more competition will accelerate innovation and drive down costs while also giving them a second source of silicon.