@Andrew Wright Thanks for your reply. That's an article someone posted in our LibreSilicon Signal group, It's in German, but you can use Google translate. https://www.golem.de/news/schneller-effizienter-guenstiger-massgeschneiderte-supercomputer-fuer-sprachmodelle-2307-175965.html They're basically doing the same I'm proposing. Our approach is, that we'll use #AI tools for place and route the MACC and ALU as hard IP cores, optimized for performance