Fujitsu's ARM-based exascale supercomputer put back by two years

Fujitsu hits challenges in shift from Oracle Sparc architecture to ARM

Fujitsu has admitted that its 1,000 petaflop ARM-powered supercomputer, dubbed Post-K, is likely to be delayed for at least a year - and could possibly be two years late.

The supercomputer, intended to be the world's first exascale computing device, has been held up as engineers get to grips with the shift in architecture from Oracle Sparc to ARM, as well as the manufacturing processes required to make the 10-nanometre parts.

The news comes just weeks after ARM unveiled further details of the Scalable Vector Extensions (SVE) technology it is developing for the ARMv8-A architecture that Fujitsu selected for the project, and which is intended to help drive ARM into high-end servers and supercomputing.

The current K supercomputer Fujitsu built for Japan's RIKEN Advanced Institute for Computational Science is ranked as the fifth most powerful supercomputer in the world, but runs customised 2GHz Oracle Sparc64 VIIIfx microprocessors, rather than ARM.

Fujitsu publicly admitted the delays yesterday, suggesting that they were due to the time it was taking the design team to get up to speed with "a new CPU-related semiconductor technology". The delay is expected to push back the switch-on date by between 12 to 24 months, the company admitted.

It is not clear whether the CPU-related problems cited by Fujitsu are as a result of the design team struggling to get to grips with ARM's nascent SVE technology (which is intended to support scalability) applying Fujitsu's Tofu interconnect ported from the K supercomputer to the Post-K design, the 10-nanometre process manufacturing of the CPUs (which will be made by TSMC), or something else entirely.

Fujitsu's Post-K supercomputer is intended by Japan and Japanese industrial policy planners to beat the US to become the first public exascale supercomputer. The US government-funded exascale supercomputer has a target date of 2023 for delivery.

Earlier this year, it was confirmed that China's Sunway TaihuLight supercomputer, featuring 10,649,600 microprocessor cores and compute power of just over 93,000 teraflops, was the world's most powerful supercomputer by a long way ahead of the second-placed design - its predecessor, the Tianhe-2.

The Sunway TaihuLight features a proprietary microprocessor, whereas its predecessor used Intel Xeon E5s.