FPGA Processor in Memory Architectures (PIMs): Overlay or Overhaul?
arxiv.org
arxiv.org
First of all, why does this sound familiar? Did early memories do something like this?
Second, how hard is it to map arbitrary complex NN to this architecture?
My worry with all these archs is the ease of debug, tracing and grokking whatever the hell is happening. Disaggregation of compute will make programs very hard to write and understand...
But it can make sense when it's a very simple and repetitive calculation. ie: copy data from one location to another, bulk bitwise ops, etc.