In my naive understanding, instead of modern SIMD type execution they were using some form of token passing between cores to execute instructions async. That made programming such multicore APUs pretty hard for the programmer. They had a SDK for mitigating some of these pain points. I could be somewhat wrong on the specifics (and the fact I last touched any architecture related topic about a decade ago).Please correct me if I am wrong.