Interesting question and I often thought about that. Optical interconnect is just a tech but likely not meant by fabric (which IMO describes the way compute units are combined). Instead we might see something like EDGE/TRIPS (which Chuck Moore worked on in the past) and Larrabee. At least this slide suggests sth like that:
Attachment 116279
Some options: smaller GPU cores spread across the CPU cores (short communication, less "reprocessing" and preparation -> higher power efficiency), compute arrays, reconfigurable computing. General purpose cores might become a collaborating set of small cores, accompanied by some bigger units for special workloads. As someone said, it might become standard to have more units on a chip (trannies are cheap) than power budget would allow to be active at the same time. Maybe I will elaborate that in more detail on my blog. No promises![]()
Bookmarks