The exascale system that will replace the powerful K supercomputer and will roll out in 2020 will use custom ARM-based chips, the OEM says.
Fujitsu will ditch SPARC chips for the ARM architecture for the next generation of the K supercomputer that currently is the fifth-fastest system in the world.
At the ISC High Performance 2016 supercomputing show this week in Frankfurt, Germany, Fujitsu officials said the successor to the current K computer (pictured)
, which is housed at the Riken Advanced Institute for Computational Science in Japan, will be an exascale system that is expected to be operational in 2020.
The future supercomputer, which officials are referring to as Post-K, will be built by Fujitsu and the government-funded Riken facility, reportedly through $1.24 billion from the Japanese government. A key part of Fujitsu's announcement is that the Post-K system will be powered by processors based on ARM's 64-bit ARMv8 architecture rather than the SPARC64 chips that currently drive the K computer.
The choice to swing to the ARM architecture was based on a broad array of factors, including the need to keep power consumption and costs down while still driving the performance needed for the massive system. ARM's low-power architecture can be found in most smartphones and tablets on the market today, and the company and its chip-making partners for the past several years have been pushing to develop systems-on-a-chip (SoCs) for data center systems, including servers, storage appliances and networking gear. They've argued that the low-power nature of the architecture can help drive down the power-per-watt of systems when compared with other architectures, including SPARC and Intel's x86 chips.
Fujitsu seems to have agreed, according to Nathan Brookwood, principal analyst with Insight 64.
"Power-per-watt matters" in exascale computing, Brookwood told eWEEK
. "You need to keep the power down. You need to get the power-per-watt down, down, down."
The decision to go with ARM is certainly a blow to Oracle, which inherited the SPARC architecture when it bought Sun Microsystems in 2010 for $7.9 billion. Industry observers expected then-CEO Larry Ellison to ditch the hardware business and keep his company's focus squarely on software assets like Java. However, Ellison kept the hardware business, building up a portfolio of tightly integrated hardware-software solutions and offering an aggressive roadmap for SPARC
Oracle also extended the SPARC development pact with Fujitsu, which had had a similar agreement with Sun before the acquisition. However, Fujitsu's decision to shift to ARM for the Post-K supercomputer will essentially end Fujitsu's development of its SPARC64 processors, industry analysts said. Fujitsu's decision also puts Oracle in a difficult position as the only vendor using the SPARC architecture in a market where Intel is dominant and ARM can point to multiple server chip partners, including Advanced Micro Devices, Qualcomm, Applied Micro and Cavium. Even IBM is looking to partners to extend the reach of its Power architecture via the OpenPower effort, though IBM is the primary developer of the chip itself. ARM creates the chip design and then licenses that design to its chip-making partners.
The SPARC architecture also is used in a segment of the server market that is shrinking as more enterprises turn to Intel's x86-based products.
"ARM's ecosystem is growing by leaps and bounds," Brookwood said. "The SPARC ecosystem is kind of stagnant. … I think one of the real downsides for SPARC is that Oracle/Sun could point to Fujitsu and say, 'See? We're not the only ones.'"
That won't be the case any longer. Ellison is pouring a lot of money into developing future versions of SPARC, but at this point, "it's only Oracle that's pushing it," he said.