Artificial intelligence (AI) reach tall efficiency operation (HPC) already was this year Xin Xianxue of industry of science and technology, japanese Fuji is connected also apply proper motion to develop special application chip in the light of AI and HPC (ASIC) , the DLU deepness that includes to learn a quantity to be made personally for AI deepness only learns special chip, and be aimed at Beijing of new generation Post (Post-K) the ARM that super computer designsFrameworkHPC chip, and two chip will by hand in byBrilliant circle generation is versed inBibcock stage accumulates report (2330) acting labour.

The depth that AI learns one of skills learns, it is the vital technology that advances AI judgement, but the biggest question that deepness study encounters at present, it is to want to undertake in huge amount material figure is handled ceaselessly. To shorten operation time, fuji is connected rolled out the deepness study servo with global the rapiddest rate implement reach operation of high in the clouds to serve Zinrai deepness to study a system, also be aimed at the character of deepness study at the same time, design DLU deepness to learn chip by oneself, predict to be expected to begin shipment this year.

relatively use plot processor at industry most (GPU) parallel operation technology will undertake AI deepness learns, the deepness of DLU much core that development of proper motion of Fuji general rule makes for deepness study only learns chip, it is the principle that uses parallel operation at the same time, but can reduce operation power comsumption effectively, the target is the deepness study chip that compares a competitor, own the every tile property of 10 times (Performance Per Watt) expression.

In addition, fuji is connected also will use DLU chip in super computer ” Beijing ” in the center, and the phase of Tofu interconnection technology that can open proper motion to design with Fuji is united in wedlock, will can build exceed extensive nerve network (Neural Networks) operationFramework.

On HPC operation layout, fuji connects as super as the Beijing of joint development of institute of Japanese manage chemistry computer, ever seized the super computer with next global operation the most powerful ability 2 years continuously 2011. Fuji is connected with Beijing super computer gives business to use the PRIMEHPC series servo that change for framework development implement, be considered the attention of the unit scientificly, this series servo implement the client beyond the first Japan is Taiwan atmosphere bureau.

What super computer basically uses the capital that Fuji opens at present is SPARC framework processor, the past is a few years appoint by the stage the generation that accumulate report is versed in, but the Post Beijing of new generation is super computer, use first degrees ARM framework to design processor. Fuji is connected use framework of ARMv8-A instruction collect, the biggest characteristic is to be able to compare other servo implement processor has lower power comsumption, let super computer when having voluminous data operation, can more effective reduce with n. As we have learned, processor of first ARM framework will use a stage to accumulate report 7 Nai riceMake ChengThe quantity is produced.