Disappointing performance results on new MacBook Pro M1 Max

[jenkinsm]

A laptop from 2021 being 20% faster than a DESKTOP from 2013 is a HUGE improvement. Compare it to a 2013 Macbook Pro and then you'll have a fair apples-to-apples comparison!

[RHenriques]

I suppose that the latest M1's are not yet there for this kind of processing. Maybe Metashape can be fine tuned for dealing with Metal.
This is the main reason why I've bought an Intel MacBook Pro 16 last January, fundamentally to be able to use external GPU's. This setup, with the internal AMD 5600 and AMD 6800 XT eGPU's is incredible fast.

[protactinium]

I just replied to an old thread with some M1 Max results, without seeing this one.  The speed-up you see will depend on the balance of GPU vs CPU tasks.  The worst relative performance is depth map generation, the best is meshing.  If your normal workload is dominated by matching and depth maps, the Apple Silicon will not seem great but for everything else it seems pretty fast.

[bgreenstone]

Yeah, that's what I saw too.  Alignment and Meshing get about a 2x speed improvement for me, but Depth Map Generation considerably less.  Overall improvement of 20% for processing an entire Chunk from beginning to end.

[jenkinsm]

I have a standard M1 Macbook Pro 13" (not M1 Pro or Max) and one thing I noticed is that generating a mesh from depth maps in 1.8 Beta FINALLY worked for me. I could never get it to work in 1.7 on Windows, the model always came out as a mushy blob.

So it may not be as fast as a desktop but at least it does some things better!

[bgreenstone]

I still get the blobs using that version on my M1 Max.  Therefore, I’m still using the old version before all of that broke. 

[jenkinsm]

I still get the blobs using that version on my M1 Max.  Therefore, I’m still using the old version before all of that broke.

Hm AFAIK only 1.8 is optimized for Apple Silicon. Maybe try that version and see if it's much faster than 1.7?

[bgreenstone]

That’s the version I tested, but I have to use 1.7.3 because it’s the last version that correctly build meshes from depth maps.  Ends up the non-native versions like 1.7.3 run about the same speed as the native 1.8.  Most of the work is done on the GPU and Rosetta 2 does a great job of converting the CPU code.

[jenkinsm]

I tried a test project on the M1 Max w/ 32 core GPU and 32 GM RAM and I was thoroughly impressed with its speed. I will do more testing to compare it to my desktop PC but I think if you're experiencing performance issues maybe there is some other problem with your project?

[bgreenstone]

It probably all depends on what you're comparing it to.  Compared to my old 2014 MacBook Pro it's *lightning* fast.  I couldn't run Metashape on that old MacBook Pro at all it was so slow.  But compared to my 2013 Mac Pro it's not significantly faster which surprised me.  The CPU intensive stuff is definitely faster, but since 95% of what Metashape does is on the GPU and that's only a tiny bit faster than what's on my old Mac Pro there isn't a huge overall gain.

[smhese]

just to reactivate this thread: I am pretty much blown away by the speed of Agisoft Metashape (1.8.2) on a M1Max - it seems to be as fast as a fully specced workstation with a full RTX2080Super GPU configuration.  My findings are here:
http://jenacopterlabs.de/?p=1572&preview_id=1572&preview_nonce=bebd602a10&preview=true&_thumbnail_id=1579

kind regards! Sören Hese

[SAV]

just to reactivate this thread: I am pretty much blown away by the speed of Agisoft Metashape (1.8.2) on a M1Max - it seems to be as fast as a fully specced workstation with a full RTX2080Super GPU configuration.  My findings are here:
http://jenacopterlabs.de/?p=1572&preview_id=1572&preview_nonce=bebd602a10&preview=true&_thumbnail_id=1579

kind regards! Sören Hese

Thanks for sharing. Very interesting results.

[smhese]

Just a tiny update to the M1Max discussion. My M1Max 10C32C arrived and I compared it directly to the M1Max 10C24C version. I found the 32C version to be another  5-10% faster depending on the Metashape process that is used. The 32C makes no difference for processes that do not use the GPU support.  Overall this was a bit disappointing. I was expecting a bit more in the range of 20% improvement but thats not what you get. A nice find however is that initial match&align is much faster when you activate "Use CPU when performing GPU acc. processing" in the Metashape GPU preferences. This speed improvement is likely a M1Max design side effect and likely also applies to the Mi1Max 24C version. With this pref setup the M1Max is even faster than a dedicated GPU like the RTX2080S or RTX3080Ti! I wonder what this means for the M1Ultra performance. However "dense point cloud processing" is slowed when this CPU/GPU pref is set.

Updated article: http://jenacopterlabs.de/?p=1572
best wishes! Sören
 

[ArnauCM]

Even if your new GPU is 5x faster there are several components that might be acting as a bottleneck. For example, the information transfer rate between RAM, GPU and CPU might be "just" 20% faster on your new Macbook Pro in comparison to your old Mac Pro, thus limiting your performance.

(I made up the numbers to make my argument. I am not saying they are real)

[smhese]

ja totally agree. I am always looking at full systems so the GPU performance might be throttled down by other components in some of the systems. But the M1Max comparison is valid never-less since the design (and speed) of the components should be (other that the additional GPU cores) the same. I get also exactly the same time stamps on both M1Max systems for tasks that use CPU only. Best, Sören