M3 Max Performance?

[vividbreeze]

I have upgraded from a MacBook M1 (not Pro) to a MacBook M3 Max.

The M1 has 8 cores, 16 GB memory.
The M3 Max has 14 cores and 30 GPU cores, 36 GB memory

I am using Davince Resolve 18.6 under MacOS Sonoma.
I am kind of disappointed of the Performance of the M3 Max.

I used a simple mov-File, length 1min, 30fps, 4k. It is me speaking in front of a backgroud. I am using the Magic Mask to mask myself.

• On my MacBook M1 it takes 1:55 min (at around 7fps) no fan-noise
• On my MacBook M3 Max in Low Power Mode it takes 1.34 min (at around 9 fps), fan jumps in after the process
• On my MacBook M3 Max in High Power Mode it takes 1.10 min (at around 12 fps), fan jumps starts in the mid of the process.

Maybe it is a difficult task to process but I kind of hoped that the M3 Max was significantly faster (more Cores, more GPUs, double RAM). Did I make a mistake?

 

[galad]

Probably Davinci needs some optimisations, or the bottleneck is not the GPU.

 

[streetfunk]

14" or 16" ?


watch this, the 14" is not made for the max chip.
it might be related or not ?...at least to the early spinning fans it might be, in case you have the 14"

 

[Chuckeee]

I have upgraded from a MacBook M1 (not Pro) to a MacBook M3 Max.

The M1 has 8 cores, 16 GB memory.
The M3 Max has 14 cores and 30 GPU cores, 36 GB memory

I am using Davince Resolve 18.6 under MacOS Sonoma.
I am kind of disappointed of the Performance of the M3 Max.

I used a simple mov-File, length 1min, 30fps, 4k. It is me speaking in front of a backgroud. I am using the Magic Mask to mask myself.

• On my MacBook M1 it takes 1:55 min (at around 7fps) no fan-noise
• On my MacBook M3 Max in Low Power Mode it takes 1.34 min (at around 9 fps), fan jumps in after the process
• On my MacBook M3 Max in High Power Mode it takes 1.10 min (at around 12 fps), fan jumps starts in the mid of the process.

Maybe it is a difficult task to process but I kind of hoped that the M3 Max was significantly faster (more Cores, more GPUs, double RAM). Did I make a mistake?

(1’55” - 1’10”)/1’55” = 39% faster

 

[vividbreeze]

14" or 16" ?


watch this, the 14" is not made for the max chip.
it might be related or not ?...at least to the early spinning fans it might be, in case you have the 14"

14". I saw the Video. But I was thinking that as the fans are not on high speed at my M1 I wouldn't even notice them in an M3 Max. I was also expecting that the process was more than double as fast due to the more than twice the mount of GPU cores and RAM.

 

[paulvee]

(1’55” - 1’10”)/1’55” = 39% faster

That was my first thought. That IS significant.

 

[vividbreeze]

Why is the significant? The Xcodebenchmark, e.g. is three times as fast on the M3 Max as compared to the M1. And with twice as much GPUs and RAM (and more PerfCores) I expected the number to by higher than 40%.

 

[galad]

It depends on how the Magic Mask filter works, is it running on the Neural Engine?

 

[paulvee]

Sounds like you’ve made up your mind already. I don’t want to engage in semantics. I have been involved with computing for many decades. Enjoy your machine

 

[vividbreeze]

HI paullvee, I appreciate your comment. I don't know how the Magic Mask filter works in detail. It is just that was expecting a higher performance increase.

 

[JustAnExpat]

HI paullvee, I appreciate your comment. I don't know how the Magic Mask filter works in detail. It is just that was expecting a higher performance increase.

40% is a great jump across two generations. Besides the jump from Intel to M1, I don't think there has ever been any huge increases generation over generation.

Another thing to think about is, are you using the free version, or Studio? The free version does not use GPU cores, so it's slower.

What's the Difference Between DaVinci Resolve Free and Studio? | Artgrid

DaVinci Resolve 17 is free, but what does the paid studio version have to offer? Let’s measure up DaVinci Resolve Free vs. Studio.

artgrid.io

 

[vividbreeze]

I am talking about the studio version. And the jump is from the plain M1 (first version) to an M3 Max. But maybe the algorithm doesn't scale linear with GPU cores and RAM as I was hoping.

 

[JustAnExpat]

I am talking about the studio version. And the jump is from the plain M1 (first version) to an M3 Max. But maybe the algorithm doesn't scale linear with GPU cores and RAM as I was hoping.

No, it doesn't scale like that.

 

[vividbreeze]

Out of curiosity - how does the algorithm work and what would it take to make it perform faster?

 

[Timpetus]

It sounds like it may be running on a single core, but I'm not an expert on the program. An almost 40% increase in single core performance is very good for two generations. If the program was properly optimized to run in parallel on all of the cores (GPU or CPU) you would definitely see a bigger jump from M1 to M3 Max.

 

[vividbreeze]

I think that the algorithm is based on machine learning and even Apple states that " .. Neural Engine is up to 60 percent faster than in the M1 family of chips ...". And this is consistent with my test.

The other benchmarks, show impressive results, but I will do other tests in my daily settings, e.g. running several VMs and Docker containers. The most annoying are the very noisy fans at high speed.

 

[Malus120]

14" or 16" ?


watch this, the 14" is not made for the max chip.
it might be related or not ?...at least to the early spinning fans it might be, in case you have the 14"

Stop.
Just because the 16" M3 Max MBP performs marginally better does not mean the 14" "is not made for the Max chip."

Does the 16" have better cooling and thus higher performance/less noise under heavy load? Yeah, because physics. This isn't new, the 16" M1/M2 Max also perform marginally better.

The 14" M3 Max is still much faster than any Mac other than the M2 Ultra & 16" M3 Max.

As for the OP's issue, as has already been noted this is probably running on the Neural Engine which is the one part that really doesn't scale from Mx -> Pro -> Max. Either that or there is some other bottleneck or optimization issue.

OP, If you want more information you might try running the powermetrics command in the terminal (sudo powermetrics) while running this task in Davince Resolve.
It should display something like
CPU Power: 1091 mW
GPU Power: 298 mW
ANE Power: 0 mW
Combined Power (CPU + GPU + ANE): 1389 mW

ANE = Apple Neural Engine, so if that's the highest then you know that's what's processing the load.
If you want something a bit easier to read you can also install this if you're comfortable in the command line

GitHub - tlkh/asitop: Perf monitoring CLI tool for Apple Silicon

Perf monitoring CLI tool for Apple Silicon. Contribute to tlkh/asitop development by creating an account on GitHub.

github.com

 

[falainber]

40% is a great jump across two generations. Besides the jump from Intel to M1, I don't think there has ever been any huge increases generation over generation.

Another thing to think about is, are you using the free version, or Studio? The free version does not use GPU cores, so it's slower.

What's the Difference Between DaVinci Resolve Free and Studio? | Artgrid

DaVinci Resolve 17 is free, but what does the paid studio version have to offer? Let’s measure up DaVinci Resolve Free vs. Studio.

artgrid.io

Intel to M1 was a jump from generation(s) old Intel CPU. Apple did not bother to use the latest Intel CPUs for their last Intel-based models.

As far as OP's case is concerned, it's not just the case of two generation jump, it's also a case of two model tiers jump. Normally one could expect both upgrades to give you 30-40% boost each but in reality it's not that simple.

 

[vigilant]

The workload itself isn’t that complex. 4k 30fps is probably bound by the dedicated video encode/decode blocks.

What else are you doing to push the system? Are you changing your workflow at all to do things you struggled to do in the past?

 

[ArkSingularity]

Intel to M1 was a jump from generation(s) old Intel CPU. Apple did not bother to use the latest Intel CPUs for their last Intel-based models.

As far as OP's case is concerned, it's not just the case of two generation jump, it's also a case of two model tiers jump. Normally one could expect both upgrades to give you 30-40% boost each but in reality it's not that simple.

If I recall, Apple did actually start putting Ice Lake chips into some of their models in mid 2020, although most of them didn't really get that facelift if I recall. It was mostly the MacBook Air and certain 13" Pro models, I'm not sure if the 16" ever got them.

(I'm a little surprised that the Ice Lake MacBook Air even got released to be honest. It was updated just a handful of months before the M1, and that particular model has since pretty much been forgotten.)

 

[theorist9]

You might want to check out this thread ( https://forum.blackmagicdesign.com/viewtopic.php?f=21&t=158630 ). According to Rohit Gupta of Blackmagic, "If your GPU is faster than ANE, it will be used. This is certainly the case with the Max and Ultra systems. There is a Preferences option to switch off ANE, so GPU is always used."

I think it would be interesting to try your comparison again while playing with these settings. That might give you a better idea of what is going on.

 

[Micka88]

One hour Da Vinci demanding video rendering tried just yesterday:

Macbook Pro 16 M1 Pro: 3 hours

Macbook Pro 16 M3 Max: 1 hour

 

[JustAnExpat]

One hour Da Vinci demanding video rendering tried just yesterday:

Macbook Pro 16 M1 Pro: 3 hours

Macbook Pro 16 M3 Max: 1 hour

Yikes! A 33% precent improvement. Very very nice!

 

[Pressure]

Yikes! A 33% precent improvement. Very very nice!

You might want to check your abacus 😅

 

[ArkSingularity]

Yikes! A 33% precent improvement. Very very nice!

I had to double check, I thought there was a Borat reference in there. 😂