Original URL: http://www.theregister.co.uk/2009/05/14/nec_pulls_out_of_keisoku/

NEC abandons Japan's 'next-gen' supercomputer

Can't afford the petaflops

By Timothy Prickett Morgan

Posted in HPC, 14th May 2009 22:19 GMT

In tough times, you have to make tough decisions, and Japanese server maker NEC this morning announced in Tokyo that it was pulling out of the Next Generation Supercomputer Project sponsored by the Japanese government. The project involved NEC, Fujitsu, and Hitachi building a hybrid scalar/vector massively parallel system.

NEC is not exiting the parallel vector supercomputer business, which at one point put the company on top of the Top 500 supercomputer list with its Earth Simulator monster. That machine - a cluster of 5,120 of NEC's SX series of vector supercomputers - was rated at 35.8 teraflops and ruled the Top 500 for the early years of this decade. Earth Simulator was a massive machine, occupying 34,000 square feet of floor space, consuming 5 megawatts of power, and costing $350m to build.

Pushing up into the 10 petaflops realm and beyond with a hybrid scalar (from Fujitsu) and vector (from NEC and Hitachi) engines was the goal of the Next Generation Supercomputer Project, which is also sometimes called Keisoku. Apparently, that's Japanese for "10 to the 16th power" (that's the kei) and "speed" (that's the soku bit) mashed together.

The Keisoku project was launched in September 2006 with a budget of 115 billion yen, about $1.2bn at the time. While NEC and Hitachi have done a lot of the design work on their side of the system, and as El Reg reported yesterday, Fujitsu has just completed its testing of a prototype of the eight-core "Venus" Sparc64 processor that was to be the scalar half of this future Japanese super, slated for initial deployment by next March and full deployment by March 2011.

The machine was to be built at Rikagaku Kenkyusho, or Riken for short, a major research institution in Japan, in a facility located in Kobe, Japan. According to a report by the Kyodo News Service, Riken is committed to now building the fastest scalar supercomputer, and Fujitsu remains committed to the project. Maybe Fujitsu will get the whole deal now?

In a statement, NEC said that it would continue to develop its vector supercomputers and was in no way shutting down or divesting this business. The company was, however, struggling financially after posting a $3bn loss in fiscal 2008, and Hitachi was slammed with an $8bn loss in its fiscal 2008. (You can read the gory details of their financials here).

Riken finalized the design parameters of the Keisoku machine in September 2007, and NEC and Hitachi have just finished up the detailed design specifications of the gear to meet the parameters for the machine. And now, while both NEC and Hitachi are hemorrhaging like pork dumplings stabbed by the chopsticks of the economic downturn, NEC got cold feet on the next phase of the project, which was manufacturing the components of the Keisoku machine.

"NEC estimated that costs for moving forward with the manufacturing phase of the Next Generation Supercomputer Project would significantly impact earnings for the fiscal year ending in March 2010, due to extensive investment required for the computer's manufacturing," the company said. "Therefore, NEC has decided not to participate in the project's manufacturing phase."

The company added that it will continue to work with the Keisoku effort in an oblique way, by adopting the new vector technologies it co-developed with Hitachi under the project and perhaps the optical interconnection technologies that were part of the system design too. NEC expects that the technologies it helped define as part of the Keisoku effort will eventually give it some leverage in the cloud computing era, but was vague - and I mean truly nebulous - about how that might happen.

Maybe Riken can subcontract the manufacturing job for the vector side of the NGSP box out to some chip and metal bending shops in Taiwan? Because Japan has so many SX and Hitachi vector applications, slotting in a Cray vector machine is not really possible because of application compatibility issues. And more importantly, the whole point of the Keisoku effort was to get the three big server makers in Japan to work together to create a hybrid machine that would allow Japan to raise its head high and top the supercomputing charts like it has in years gone by.

And to do a lot of research, of course. But with those kinds of funds, the politics matters as much as the research does. And that NEC and Hitachi can't follow through on the Keisoku effort after the hard design work has been done is embarrassing to the Japanese government. It will be interesting to see how the government and Riken respond. ®