Quick thoughts after Cambricon-U.
RIM shows our perspective of PIM as researchers who did not studied in the field. When it comes to the concept of PIM, people intuitively first thought of the von Neumann bottleneck, and believe that PIM is of course the only way to solve the bottleneck. In addition to amateur CS enthusiasts, there are also professional scholars who write various materials in this caliber.
This is not true.
In RIM’s journey, we showed why previous PIM devices could not do counting. Counting (1-ary successor function) is the primitive operation that defined Peano arithmetic, and is the very basis of arithmetic. We implemented the counting function in PIM for the first time by using a numeral representation that is unfamiliar to scholars in the field. But this kind of trick can only be done once, and the nature will not allow us for a second time, because it is proven that no representation can efficiently implement both addition and multiplication*.
Computation implies flowing of information, while PIM wants them to stay in place. Contradictory. Combined with Mr. Yao’s conclusion, we may not be able to efficiently complete all the basic operations in-place in storage. If I were to predict, PIM would never really have anything to do with the von Neumann bottleneck.
* Andrew C. Yao. 1981. The entropic limitations on VLSI computations (Extended Abstract). STOC’81. [DOI]