Hi @kjetilpettersso​ ,

Thanks for sharing your testing and results with FME on the M1 Apple Silicon Mac.

Our team is aware and is working on doing some testing ourselves. Our current plans regarding development for native support on Apple Silicon are still not determined but it's on our radar.

I'd recommend signing up for our FME Explorer's program and we'd be happy to reach out to you if we undergo any user testing/experience.

- Andrea

Thank you. I will do that.

Over the break I had a chance to try an M1 Macbook Air. I was using FME 2021 betas (and ironically an overlay as well). You can see my tweet about my experience over at https://twitter.com/DaleAtSafe/status/1344439793854234624, but in short, I found quite comparable performance between intel mac and m1 mac doing a massive point on area overlay (which seems to be everyone's favorite speed test case...)

Now, in my case I have feature table/bulk mode input — I'm reading CSV and Shape — and in FME 2021 we've greatly optimized memory use in that scenario. As in on Intel anyway it uses 1/7th the memory, at least in this scenario, as compared to raw features.

I looked at your test case and if you wanted, you could put in AttributeKeepers and turn on the "create bulk mode" switch if you wanted to try FME 2021 beta on that. Or recreate your FFS files to be feature tables using this trick and then run from there -- that would give best performance.

I too saw that the memory reported was greater on the M1 Mac than the Intel Macs. but saw very comparable performance. It could be the way the architecture reports things, or it could be because we're not native, but agreed that it seems the base memory overhead is higher on M1...but if we can use bulk mode we're set up for the win it seems.

Will be very interesting to do more testing and understand the scenarios better.

In terms of porting FME, our biggest hurdle is that we use Qt, and officially we've read that Qt won't support M1 until Qt 6.1. They just released 6.0. But we do see this as an important opportunity and are pretty excited to see what kind of performance we'll end up with once we're over on that side...

I ran the unaltered workspace on FME2021 Beta (Build 21258) and it finished in 5:59 (down from 8:41). FME still logged several memory optimization messages, but no errors this time.

Using the "attribute keeper/create bulk mode"-trick further reduced to the running time to 3:52 (!) and no memory optimization messages whatsoever. I need to remember to use that more :)

Since Safe have released an Apple Silicon native version of FME Workbench 2023 I decided to run the same point-on-area overlay analysis benchmarks (more or less) again. For comparison I also ran the same benchmark on FME Workbench 2021 (using Rosetta) and on both FME versions running on Windows.

I used these two systems:

MacBook M1 Pro 16 2021/32 GB RAM/MacOS 13 Ventura

Dell Precision 5530 Core i7-8550H/32 GB RAM/Windows 10

The results are quite impressive:

2.7M points/1.9M polygons

FME 2021, MacOS Rosetta: 02:47

FME 2023, MacOS ARM native: 00:47

FME 2021, Windows: 04:24

FME 2023, Windows: 02:37

4.8M points/2.8M polygons

FME 2021, MacOS Rosetta: 04:19

FME 2023, MacOS ARM native: 01:16

FME 2021, Windows: 04:49

FME 2023, Windows: 04:34

For mor details read the blog post.

Thanks for sharing! My current Mac is still going strong but as soon as it starts showing signs of kicking the bucket I know what I'm getting 😁