New results show it's 5x slower on the M1 https://medium.com/analytics-vidhya/m1-mac-mini-scores-highe...
Interestingly, it's only utilizing the neural engine and no GPU. It'd be interesting to see what happens when M1 Tensorflow updates to utilize both GPU + NE
It's like you're moving just across the street, and loading every single box into a car, crossing the street, then unloading the box instead of just walking on foot. You need to drive further (bigger networks) and load more boxes at once (batch size) for a car to actually be useful in this scenario.
To facilitate Nvidia's tensor cores OP had to use Nvidia's own TF distr./image and configure it explicitly. Something PyTorch does out of the box. Nobody knows why Google doesn't do this, maybe they want to push their own Cloud TPUs.
> Adding PyTorch support would be high on my list.
Won't happen. PyTorch needs Apple's help bc of the lack of docs, they've asked already and Apple hasn't commented or promised any kind of support, nothing. That they've chose TF instead of the current market leader doesn't give me too much hope and might come from backroom deals we don't know of.
Wondering why OP didn't invest the money into a 2nd 2080 Ti.
- my arduino scores higher than my raspberry pi in gpio speed tests
- my honda scores higher than my peterbilt in pizza delivery test
I picked up the baseline model for $699 (8 GB unified RAM, 8/8/16 cores version with 256 GB of storage) and I've mostly been tinkering, as I use Windows 10 Pro / Fedora 33 on my workstation / gaming computer, but I can tell you it runs World of Warcraft: Shadowlands at 60 FPS 1440p resolution with Ultra settings, which is pretty damn astounding... I mean, I know WOW is running on a nearly 20 year old engine, but that engine has seen a hell of a lot of refinement over the years, and many advanced features have been added.
The M1 is a testament to Apple's engineering team. I really look forward to seeing what they could do if went buck wild and gave themselves a 95 - 180 watt TPU range to compete with Ryzen 5000 / Threadripper 3000 series parts.
We've seen that the M1 is competitive in low power scenarios... I want to know if it can be scaled up and be competitive when power is no concern.