M3 Pro benchmarks vs real-world usage

[bcortens]

Why openCL instead of Metal? Not sure if an openCL benchmark has much relevance.

Mostly because there weren’t many benchmarks in metal of the unbinned chip.

The M3 pro is lower in the lineup than the m2 pro was (that is to say there was more performance distance between the m2 and m2 pro than between the m3 and m3 pro).
Repositioning the m3 pro down a little isn’t a bad thing, I find it disappointing because I thought the m2 pro was practically the perfect balance. Apparently too perfect.

 

[JinxVi]

In any case we do know the single core is much better and that alone is worth it to some.

But single core performance is a reason to buy an M3, not an M3 Pro. The Pro needs to justify its existence in improved multi core performance.

 

[Mcdevidr]

But single core performance is a reason to buy an M3, not an M3 Pro. The Pro needs to justify its existence in improved multi core performance.

Well yea but not when they are basically the same price if you want more ram than 8gb. Anyways m3 pro is faster than m2 pro in multi core as well just not much. In any case this has been discussed a bunch. I ain’t really got much to add.

 

[leman]

Lets go to GPU:
The M3 Pro is, on average, slower in metal compute than M2 Pro M2 Pro vs M3 Pro GPU Compute

Unfortunately, the link you provided does not open for me.

Quantifying GPU performance between M3 Pro and M2 Pro is a bit tricky. On Geekbench they are about the same overall (M2 Pro can be few % faster). The GPU runs at the same clock frequency and the basic shader compute layout did not change, and M3 Pro has fewer GPU cores on average. Also, M3 Pro has reduced the memory bandwidth by whopping 25%. All these things do put M3 Pro at a disadvantage. At the same time, M3 Pro GPU is capable of superscalar execution and generally runs shaders more efficiently. The more complex the shader, the better M3 Pro will perform in comparison.

 

[name99]

so $100 per P-core

Of course you're actually getting SIX more P cores (and losing two E cores).
But details never stopped anyone...

The more interesting thing about the Max is the number of transistors.
Substantially more than 2x the transistors, but only barely (and with some handwaving) 2x the CPUs and GPU cores.
Where are all those transistors going?

Even if you don't care about many displays, presumably they are spread over more memory controllers, more SLC, more NoC queueing, more IO, just more more more stuff that is hard to isolate in benchmarks or to talk about in ads, but which continually feels like the machine is just not slowing down in a way that a Pro might when pushed?

It does look to me like, in the M3 generation, the lower-end M3 Max is probably the bargain of the line for people buying a high end mac.

 

[name99]

You can see a different set/type of benchmarks here:

Comparing Accelerate performance on Apple silicon and Intel cores

Comparison between 2 Intel and 2 Apple silicon Macs running vector and matrix functions from Apple’s Accelerate library. Was that new M3 worth the money?

eclecticlight.co

As usual Howard tests small matrices (like 16x16) which I'm not wild about because I think you get swamped by overhead. Of course for some people that's the case they care about, and so they want to test that overhead, but it's definitely not going to show AMX to advantage.

For this sort of task the M3 Pro is clearly better than Intel 8 core, while being available much more in the low/mid range, not only the high end.

I'm still looking for M3 Pro/Max benchmarks of AMX that test the system at its optimum (eg large matrix multiplies). It's still unclear to me if the M3 Pro (6 P-core cluster) has one or two AMX units...

 

[pshufd]

When I look at my M1 Max cores (work computer), the 2 efficiency cores are being used all the time while the 8 performance cores are mostly idling. I am running docker containers and compiling rust code in the background and yet still my efficiency cores are doing most of the work.

So I guess for my usage (coding) I would really benefit from having more efficiency cores like in the M3 Pro (which also have improved by 20%), especially for all the small services and apps running in the background (chat apps, email etc) during a normal work day.

Yes the benchmark scores are the same as M2 Pro, but I think having 2 more and improved efficiency cores has a greater impact on my kind of day to day work. But maybe I am wrong?!

All the negative comments around the M3 Pro almost pushed me towards a M3 Max for my private MacBook purchase. Would be great to have some more tests from the M3 Pro...

This is what my M1 Pro MacBook Pro use looks like too.

My production use has high program startup usage and then it's just mostly the efficiency cores running during the day. I suspect that I could run completely on efficiency cores though startup time would take longer.

It's nice to look at what Apple is doing but I have no need for a new laptop. At some point, Apple will make it worth my while to upgrade, but short of loss, theft or destruction; I usually use my MacBook Pros for a decade and typically upgrade after 7 years.

 

[Kotsos81]

I had an Intel MBA from 2014. I develop optimization algorithms and run simulations in Matlab. The MBA needed about 10 hours for a particular simulation. I recently purchased a MBP M3 Pro 12C with 36GB of memory. Using Matlab R23b native for Apple Silicon and MOSEK 10.1 toolbox native for AS as well, it took less than 3 minutes for the same simulation!

Well, needless to say, that's fast enough for me!

 

[dgdosen]

Interesting? MPB analysis linked on Hacker News (FWIW):

Tracking developer build times to decide if the M3 MacBook is worth upgrading

When our CTO said "I'll upgrade your MacBook if you can prove it's worthwhile" we embarked on a journey including (re)building a Go hot-reloader, instrumenting developer builds, analysing compiler performance and feeding an AI model the data until we had an answer.

incident.io