OPF eyes more Power innovations, partners

Using the Power architecture, the Open Power Foundation (OPF) is pushing for computer innovations ranging from supercomputers to flash memory.

The OPF did fundamental work on the Coral program, a supercomputer architecture for the U.S. Department of Energy. The group also aims to pioneer memory capabilities for servers based on flash, while partners in China and elsewhere are gearing up other innovations.

The Coral program will forge a Power 8-based supercomputer architecture in collaboration with the Oak Ridge, Argonne and Livermore U.S. national labs by 2017. It will use accelerators such as graphics chips linked over Nvidia's NVLink and other chips riding IBM's coherent accelerated processor interface (CAPI).

The government labs "need a bigger scale than most anyone else in the world," said Brad McCredie, vice president of IBM Power Systems Development and OPF president. "That scale drives so much hardware and software interaction... reliability, efficiency and systems clustering connectivity are the biggest challenges we're going to face."

IBM won the contract in part because it could offer a ready-made ecosystem with partners such as Nvidia, Mellanox and the OPF, McCredie said.

"While [national labs] do need to drive the pace of technology innovation and need some really, really big systems with very, very large capabilities, they are very heavily focused on trying to keep needs in line with the market. Like everyone else, they need to have ecosystems around them in order to draw on to build technologies they want," he added.

On the memory front, McCredie said the foundation's work in low-latency flash storage will be disruptive. At the chip level, it is helping define ways to plug flash into DRAM and CAPI interconnects to enable new capabilities in big data analytics software.

At the software level, the OPF has developed BigInsights, its own version of Hadoop for processing big data. Fadi Gebara, a senior manager at IBM's Austin Research Lab, said it will also use Java on GPUs to help visualise analytics.

Working with the start-up Diablo, IBM put flash on DRAM DIMM modules to deliver new capabilities. "It allows you to buffer memory right into DRAM and then execute into a flash storage module that sits on a memory bus," Gebara said.

The Diablo technology is well suited for applications such as high-frequency trading, where low latency is paramount, he said. IBM is also working on applications for flash memory riding its CAPI interconnect.

Playing the China card

IBM expects its collaborations around Power processors to expand, especially in China.

China's Suzhou is in the last stages of finalising agreements to build Power chips. An event in China brought "lots of support to start building ecosystems around the Power chips that Suzhou is building," McCredie said.

1???2??