This is a great infrastructure improvement. If you have a two lane expressway supporting a neighborhood, widening the road will reduce bottlenecks and improve speed. But, no matter how wide you make that expressway, coming into and going out of each house is limited by a driveway. And most of the time, you cannot come in and go out at the same time. This. This is the restriction of von Neumann architecture. So, no matter how you improve the interconnect, there are "driveways" in and out of each core within the most powerful AI processing chips. Whether those cores are GPU, TPU, CPU, or any other PU.
This is the benefit that a migration to compute-in-memory from Von Neumann offers. I mean a true compute "in-memory". Not a von Neumann wrapped in a PIM or CIM name. Then optical interconnected freeways can provide even greater benefit by only bringing in the new information that needs to be processed, not moving around the data being processed. Hello neighbor, can I borrow a cup of sugar to make my pie?
Ayar Labs aims to reduce AI infrastructure bottlenecks using optical innovations. CEO Mark Wade discusses their latest $155M funding round and how they're working with partners like Nvidia, Intel and AMD https://trib.al/nHxUk3O
Investor
3wCongrats to Ayar Labs and Sivers Semiconductors AB this is the future!