Apple's work is cut out for them; M7 and M8 desperately need a better GPU. These SOCs put up fantastically efficient single-core performance, but have some of the worst GPU compute efficiency you can buy. It also needs a better interconnect than RDMA, since Thunderbolt is snail-speed when networking several iGPUs together.

Apple Silicon is a mirror into the Mac's decision paralysis. Apple wants to unify the Mac power users and iPhone casuals onto one silicon backbone, but can't decide who gets the short end of the stick. Tim Cook especially couldn't figure out how to ship a Mac Pro while retaining "ecosystem" level control over academic and prosumer customers. Professionals and researchers had a stack they wanted to use, and Apple didn't support it.

Ternus will probably try to pivot Apple Silicon towards professionals, but he'll need to make software concessions to enter the true ARM HPC market. Ironically, that may entail Apple's rapprochement with Khronos to get the industry onboard with a CUDA stack alternative. Even a basic "inference hardware" GPGPU standard would put Apple Silicon back on the map for professional applications it has otherwise sat-out.