The Next 3x in Inference Won't Come from Faster Kernels (twitter.com)

One unique idea for solving the chip shortage: pack more models into each existing GPU by having them share the resources. :-)