Making ten Peta flops peak machine is making one Exa flops peak machine -- Is it hard or easy to make and to use?

Yasumasa Kanada

Abstract

RIKEN, Fujitsu, NEC and Hitachi released a news on the 14-th of September, 2007 as follows; Japanese project of making sustained performance of 10 Peta flops machine with Linpack will be completed by the end of the fiscal 2012. Operation will start by the end of the fiscal 2010, i.e. two years time from now and another two years for the enhancements focused to the required performance. Total budget needed for constructing the machine including software (OS, compile, library and application) and building for it is 115 Billion Yen (about US $ 1.3 Billion.) The machine is completely heterogeneous and consists of general multi-core CPU based less-power-consumption scalar part and special processor based less-power-consumption vector part. 45nm process semiconductor technology must be used for the chips. Optical interconnect technology should also be used. Then the system should be compact. -- Detailed system configuration or architectures are not cleared or opened to public since then.

For constructing and utilizing very fast machine, we have to take care of system availability, coverage of the application fields, correctness of the computation and real performance. In my talk I will mainly focus on 1) what will be the key issue for making and using the Exa flops machine, 2) how hard or easy to construct and utilize the machine. Keywords : Low voltage, Signal Margin, MTBF, Power Supply, Pins for signal transmission and electric power supply, Reliability, Availability, Correctness of the computation.

back to agenda