Sure, as a badass SSH machine, it works great for deep learning... when SSH'd to a server with NVIDIA cards. Now don't get me wrong, I absolutely adore my M1 M

Sure, as a badass SSH machine, it works great for deep learning... when SSH'd to a server with NVIDIA cards. Now don't get me wrong, I absolutely adore my M1 MacBook Air. It's so good that it I've kept my i7+2080+32GB RAM desktop turned off for weeks now (outside of gaming). My 8GB M1 MacBook Air is now my main work computer, and it's spectacular. That said, the basic MNIST timing benchmarks are decent, but don't get your hopes up for anything more anytime soon. There's simply no way in hell today's M1 SoC can train much, much bigger & complex models than hardcore systems with beefy GPUs with 12+ GB RAM on each card. So no, today, the M1 MacBook Airs/Pros are not really good for DL. And honestly, that's fine. Maybe Apple will compete with NVIDIA down the road, and I'd be happy to see that. But I'm glad the author came to the (correct) conclusion that M1s just aren't anywhere near there yet for anything more than basic DL. Edit: that all said, there might be an interesting use case for M1 Mac Minis as edge inferencing nodes.