>> "The fixed function nature of [GPU] shader cores makes it difficult to run non-gaming workloads on them. This meant that game technology innovations were often held in check by graphics hardware capabilities."
Sentence 1 is strictly untrue, else NVidia wouldn't be invading the scientific/engineering/AI and data mining markets so heavily, and whether true or false, setence 2 doesn't follow from sentence 1.
>> " One of the key drawbacks to the ring topology is that data needs to pass through each node on its way. The more cores you have, the greater the delay."
It wasn't really the ring topology so much as the exponential communication requirements of MIMD architectures. Intel made the same mistake nearly twenty years before Larrabee, in it's "Personal Supercomputer" project. It used a hypercube-based topology, which avoided the delays in the ring structure used by Larrabee, but still failed due to the massive intercommunication bandwidth necessary to feed those processors.