Comparing Apple M1 to RTX 3060 to AMD 5900HX CPU to AWS buck an hour GPU

I rented a little Mac Mini for the day as part of my project (it’s called Lartis now)

I wanted to get a Mac M1 version of Lartis working (Bad news there - it’ll cost me way more than I can afford ATM to do a Mac version - just discovered this…)

Anyway - the figures are interesting - for those who want the TLDR the M1’s not a repacement for a decent GPU.

So, here’s what I’m aiming for - the end result…

This image will render at a decent speed (i.e. a number of seconds) on anything. However, creating the style in ther first place is a very computationally expensive task. So, I’ve been collecting some timings by making the same model several times.

On my 3060 GPU the style takes 1 nour and 45 mins to create (so if you’ve got a decent GPU you’ll either be faster or a little slower than me)

If I ignore the GPU and just throw all the CPU power I can muster at creating exactly ther same model it’ll take about 36 hours (note I didn’t actually let this one finish - Lartis gives me an ETA so I let it run for a few mins then aborted the job)

The Mac M1 does a respectable job about 8.5 hours - this is not exactly what I was expecting. Everyone always goes on about how fast they are… Having said this I am using the very most basic M1 you can get - a M1 Mini with 8Gb of RAM - more memory would definitely help as would a better CPU as found in the more expensive Mac M1s

Top prize goes to Amazon which blows everything else out of the water at 42 minutes. This run was on a AWS server. The server has 16 Gb of RAM (not that I use it) and a decent enough CPU. What makes all the difference is the super fast A10G GPU it has which includes 24Gb of GPU RAM. This is about the speed a 3090TI can manage according to some benchmark figures I’ve read but those things cost serious cash while the AWS server costs a buck an hour… (1 / 1800th the cost of a 3090TI)

Think I’ll go “invade Russia” to see how the Russian Cloud PC compares later… It’ll give me a good comparison to my 3060 and be about the speed of a 1050 or 60 (somewhere in between actually)

Oh - M1 just finished the training - accurate figures…

Done, trained model saved at models/mac-test-256.pth
Batch size = 5 - Epochs = 1
Elapsed time = 31035.562179 secs
Elapsed time = 8 hours 37 mins 15 secs

I tried replicating the above image on the M1 with DAE but DAE crashes on M1s (not tried this before)

Here’s one from Windows to make up for it

Dunno about you - I like the Lartis version better (zoomed styles… yum)