China defence industry sign deal for an new neural-network chip: Han-Wu-Ji

[52051]

China PLA committe for science and tech signed 1.35 billion RMB (roughly 200 million USD) contract with insitute of computing science from China academy of sciences, for the latter to develop neural networking computing processors for military usage (most likely computing vision etc) based on Hanwuji ISA
(more detail about this neural-network CPU can be found here (in Chinese): https://zhuanlan.zhihu.com/p/23255679).

China's hedge funds will also add another 1 billion RMB (roughly 150 million USD) investment for this chip set project.

It is reported that this is the largest contract the insitute of computing science have ever signed.

https://lt.cjdby.net/forum.php?mod=viewthread&tid=2399499&extra=page=1

 

[52051]

This is basically my area, this CPU structure is promising, it is like a GPU, but with much faster/larger shared memory and much more efficient FMA and matrix ops codes.

And unlike Nvidia's GPU and Google's TPU, Hanwuji dont use register, instead, their shared memory is at the speed of register and use as register for vector computation, basically they give the progammer a huge size of address-able/programmable register-speed memory, thats a huge advantage over existing program model.

It is more general-purpose comparing to Google's TPU, and more energy-efficient than Nvidia's latest deep-learning GPUs (volta), and comparing to Intel's rather awkward solutations like xeon phi and skylake Scalable, it has far better energy efficiency and will take far less effects to writing highly efficient computing code thanks to its programmable and addressable "register" array and vectorized programming model.

And this CPU structure is very scaleable as well, it is said that China's Huawei will liscence this Hanwuji ISA to their next gen mobile chip for phones.