Gb now has reports of base M3 hitting over 3150. How is this not a good ST score? That's the same kind of improvement as we had from A13 to A14.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Are we sure M3 is based on A17 Pro?
- Thread starter senttoschool
- Start date
- Sort by reaction score
Gb now has reports of base M3 hitting over 3150. How is this not a good ST score? That's the same kind of improvement as we had from A13 to A14.
Simple:
A15 --> A17 Pro: 26.64% improvement in ST
A15 --> A16: 11.6% improvement in ST
M2 Max in a MBP --> M3 Max in a MBP: 10.7% improvement in ST
How is that good? You tell me.
At first glance, doesn't it look more like M3 is based on A16? The improvement is within A15 --> A16 range, not A15 --> A17 Pro.
Where is the 26.64% improvement in ST that we should have seen from M2 Max to M3 Max? Heck, let's be conservative by saying scaling isn't as easy anymore. But it didn't even hit 20% improvement.
Source:
A15, A17 scores: https://browser.geekbench.com/ios-benchmarks
M2 Max ST Score: https://browser.geekbench.com/v6/cpu/3354195
M3 Max ST Score: https://browser.geekbench.com/v6/cpu/3364975
Last edited:
I'm guessing you're the one that just posted these on reddit too?
Anyway (and not sure how "proper" this is but), given most of the ST uplift has been through frequency increases we can just look at those.
A15 (3.2GHz) -> A17 (3.8GHz): 19% frequency increase
A15 -> A16 (3.5GHz): 9% frequency increase
M2 Max (3.8 GHz) -> M3 Max (4.05 GHz): 6.5% frequency increase
However if we look at the M2 Pro -> M3 Max in terms of highest scores (and assuming the higher M3 Max scores I'm seeing are background processes being done), we get 2686 -> 3227 which is 20% ST improvement. Not quite 26.6% but closer than 10.7%. That's what I think anyway.
That's one way to look at these things. Let's try another one (I am only looking at highest scores I can find in GB browser, not the published median, for obvious reasons)
Code:
A15: 3.2Ghz GB6 2400
A16: 3.5Ghz (+0.3) GB6 2650 (+250)
A17: 3.8Ghz (+0.3) GB6 2950 (+300)
M1: 3.2Ghz GB6 2400
M2: 3.5Ghz (+0.3) GB6 2650 (+250)
M3: 4.5Ghz (+0.5) GB6 3160 (+460)As you can hopefully see, it looks like M3 has "jumped over" a generation. The improvement from M2 to M3 is almost doubled compared to improvement from M1 to M2.
Two comments: first, I think we should be looking at absolute instead of relative improvements, as there is good evidence that this is what Apple is actually pursuing. For a few years, the numbers aligned in a way that their +250 points per generation pretty much equalled 20%, but as the performance is higher now, the relative difference is smaller. Insisting on 20% every generation means that you expect them to make bigger and bigger improvements each time, which is not realistic (in fact, I'd expect the improvement rate to slow down over time). Second, I don't think that M2 Max is a good example because so far it was the only model clocked higher than the rest of its family. We don't know whether M3 Max will follow the same route (maybe in high power mode, or maybe Apple is leaving that to the desktop versions). For this reason I am comparing the base M3 to M2.
A17 Pro’s Neural Engine is mainly to support the IPhone 15 Pro’s camera improvements.
A MacBook Pro wouldn’t need this and its current M3 NPU integration is definitely more than enough because of the sheer amount of RAM.
This balances out the die size and costs as well, so I get what you’re saying.
A MacBook Pro wouldn’t need this and its current M3 NPU integration is definitely more than enough because of the sheer amount of RAM.
This balances out the die size and costs as well, so I get what you’re saying.
Maybe it’s just me but the ST performance improvements don’t get me all that excited. Not because it’s not important, but it’s an isolated test compared to the aggregate of the various cores compared ST.
The lift on performance should have a fairly clear “more cores, more speed” to it that isn’t well demonstrated in ST.
I hear what you are saying, and I agree… to an extent. The proliferation of ML development though would stand to benefit from higher performance NPUs.
Sure, it’s not as “exciting” as photography, but this is a key demographic that shouldn’t be ignored.
Apple has probably given the best 3nm node to the iPhone and the lesser 3nm node to the Mac.
That probably explains how Apple is able to produce other 3nm chip at the same time next to the A17 Pro, with the highest quality one reserved for the iPhone.
What jumped over a generation of performance?
I'm willing to bet if Apple allowed us to overclock the M1 Max from 3.2 ghz to 4.1ghz, the M1 Max would probably be doing similar speed as the M3 Max in terms of singe core performance.
95% of models are running on GPU/CPU or/and optimised for it, folks in ML spaces don't have time to fight Apple and their closed source API's for NPU. (Some models works fine with CoreML and NPU but its minority and also Coreml is very limited).
I’m fairly certain that a lot of the media applications like Affinity Design, and Photomator use the NPUs. What’s not necessarily clear though is if these workloads scale with the number of NPUs to give you a consistent benefit.
You can check if they are using NPU on your system by simple turning on powermetrics in terminal and watching ANE power usage.
Code:
sudo powermetrics -i 500They can possibly use it especially with smaller models and smaller img/video sizes but as I said we don't have any proper API's for normal usage so you are limited to CoreML or MPSGraph (which can't guarantee usage of NPU especially with custom operations, so you still have to think about GPU performance). ANE is pretty powerful and cool but software is lacking.
The NPU has a very specific function — running simpler ML models at very low power. It is not intended as a general ML inference solution and it's not something you target your cutting-edge models on. That's also why a fast NPU is more important for the iPhone than for a Mac.
As long as Apple can use it for Siri, image classification, computational photography etc., it fulfils its intended purpose. And if you have an application that needs to run a similar type of ML model and you want to do it without impacting the battery, maybe it can be useful for you too. Different tools for different goals.
Tbh it would be very useful for ton of models even larger especially on lower-tier Macs (M1 NPU vs GPU is like 5-6x performance) its just Apple software division that are making it is being disallowed to share internal API's for NPU programming or just don't care enough / haven't enough budget and that is punishing external developers by disallowing proper NPU programming. (which in a way killed a lot of hype around ANE after M1 release)
Best example is https://github.com/apple/ml-stable-diffusion it was late to the party by about a 6-7 months but still pretty impressive stuff with very fast inference on ANE.
Last edited:
This was a great reading and seems like the most plausible scenario.
Apple M3 chips appears to be a hybrid of the A17 Pro, A16 Bionic designs
The new M3 chips are here and there’s a lot to digest. As Apple doesn’t make a point of revealing the design origins of its chips, it can be an interesting exercise to put the pieces of the puzzle together.www.notebookcheck.net
CPU already seems fishy which might be A16 based.
Which means we have about ~20% increase in ST performance to look forward to for the M4. As long as they don’t go for prolonged battery life/lower freq and power draw vs M3, which I hope they don’t.
GPU performance for M4 should be similar though, since ’original A16’ aka M3 has the same GPU as A17 Pro aka M4.
How is this a plausible scenario? We know for a fact that the CPU cores of M3 and A17 are the same and very different from A16.
That is not what the article says. If I have missed evidence of them being the same, please share / point me.
How about this:
If you look at an A17 die shot and can spot the P-cores, they look very nearly identical to M3 P-cores. I mean, N3 is a significantly different process than N4. It would make zero sense for Apple to try to port a N4 design to N3.
N3B is the only active node at TSMC (other than some larger node jobs). Samsung has a node they call N3 (or 3nm), but it is vastly different from TSMC's N3, and Apple is not contracting wafers from Samsung. N3E is not in production at TSMC yet, so Apple is not getting chips burned on that node yet. TSMC has been producing something on the order of 40 wafers a day since March, which is around a thousand good chips a day. They have probably been getting M3 SoCs since June or July, alongside A17s.
Last edited:
I dont get your point.
1) I dont have a die shot of the A17 Pro.
2) There would be no back porting. The original A16 that was destined for iPhone 14 Pro was 3nm, but had to be ported to 4 nm and gimped of its GPU when TSMC delayed 3nm. This is according to the article at least.
Edit: the macrumors reply function on mobile is written by an idiot, sorry for the formatting.
- Apple uses the same exact phrases to describe the both (wider execution, improved branch prediction)
- They look very similar on the die photos (and different from A16 cores)
- They have the same IPC in Geekbench (and different from A16)
- People who did microbenchmarks (like Geekerwan) say they are the same
P.S. Here are the P-cores from die photos, extracted by @altaic (who did an amazing job I might add!)
P.P.S. At the same time, there is remarkable similarity in the basic block layout between A16 and A17/M3, which is different from M1/M2. This again highlights the status of the A16 as transitional technology.
I’d like to see those benchmarks by geekerwan. I tried googling it but couldnt find it, please link me.- Apple uses the same exact phrases to describe the both (wider execution, improved branch prediction)
- They look very similar on the die photos (and different from A16 cores)
- They have the same IPC in Geekbench (and different from A16)
- People who did microbenchmarks (like Geekerwan) say they are the same
P.S. Here are the P-cores from die photos, extracted by @altaic (who did an amazing job I might add!)
P.P.S. At the same time, there is remarkable similarity in the basic block layout between A16 and A17/M3, which is different from M1/M2. This again highlights the status of the A16 as transitional technology.
He seems very thorough, unfortunately I dont understand a single word. I saw some charts in english when I scrolled quickly. Will try to make out some of it later, thanks!