Von Neumann may well be rolling over in his grave! Instead of constantly improving the bandwidth of the memories that feed our insatiable processors, why not reverse the problem and use memory not only as a simple space storage but also as an active processing provider? This is what Micron, who also invented the Hybrid Memory Cube, did with the Automata Processor. This literally revolutionary concept has the potential, at least on paper, to compete with coprocessors and accelerators in massively parallel computing.
What does it look like? Imagine an FPGA containing a grid of hundreds or even thousands of programmable computing units which, unlike a GPU running the same operation in parallel on Gigabytes of data, would be able to discretely parallelize different operations. The programming of such a component, i.e. the transformation of a given application into a dedicated machine, will not, however, be child’s play. To facilitate it, Micron is currently working on an SDK that should be available before the summer. We’ll keep following the story as it develops. Stay tuned!
© HPC Today 2021 - All rights reserved.
Thank you for reading HPC Today.