Recent Apple silicone with unified memory seems to have a strategic advantage over a traditional CPU/GPU pairing.

E.g. M2 Ultra with 192Gb RAM which can be allocated to any activity like ML. And the M3 series is apparently just pushing on this track.

Anyone with concrete feedback on all this?

Scientist and engineer interested in Life.

Work on (GOF)AI, Agriculture, Biology, Energy, Robotics, Space---all topics climate-aware.

And asking to smart and trusted community here, as the real question is about, well, software support—difficult to get solid feedback without Warpcast.

Personal context is that one finger on ordering a pricey Nvidia-based shared machine, while I am basically giving up on Apple. But as a shared machine, an M3 may be faster and even more cost effective?

It really is so still today, close to 10 years after Tensorflow opening. There are many AI chip contenders (each with SDKs) and innovation (=new APIs), yet little surprising many need stability on the software side.

On the other end Apple has now significant volumes (on the issue at hand). Strange attractor soon?

NVIDIA and Linux are the more travelled path, it will be much easier to get public research code etc working rather than having to in many cases port code that assumes NVIDIA hardware or directly depends on CUDA.