Fujitsu releases details of new AI processor

Fujitsu, the makers of the most advanced CPUs for supercomputers and SPARC servers, has decided to give Intel, Nvidia and AMD some competition in the AI and deep learning department.

The DLU(deep learning unit) coprocessor consists of 16 DPE(deep learning processing elements), each of which contains eight SIMD execution units, specifically tailored for deep learning workloads. 
It includes on package HBM2 and on die 6 dimensional Tofu interconnect, which will let it scale massively to tens of thousands of nodes, like PrimeHPC does.

A Deep Learning Unit with six processing units will contain for example 96 processing elements. Fujitsu also talks about deep learning integer.

The company claims to be able to achieve the required precision with 8- or 16-bit data sizes , which right now is just for full-FP32 calculations, and that is a massive time saver, it should be just as accurate. 

They are targeting 10x the performance of "the competetion", and since its release is schedueled for FY2018, that means the latest Nvidia and Intel deep learning chips.  
Given their history with K Computer, and the fact that its still at the #1 spot in HPCG and Graph500, I am inclined to believe that they will hit 10x their competitors performance.

Thanks Exascale for this spot.