So it looks like there are going to be some big changes coming to Apple and this could signal that the relatively slow updates to the Mac line are ending (and don’t get me started on the dog breakfast that is the iPad). But at the top level, the way to think about this is that there are three (OK four really different product lines), and with Apple Silicon, they seem to be running very different cadences and also they are in the middle of trying to differentiate what they are doing into “good, better, best (and ultimate although the names are confusing in terms of actual products, so one tabular way to look at what they have is:
| Product | Chip: Cadence | Good | Better | Best/Ultimate |
| iPhone | A-Series 12 mo | SE 3 A15/24 mo. | iPhone 15 A16 | iPhone 15 Pro/Max A17 Pro |
| MacBook | M-Series 18 mo? | MacBook Air M1, M2 | MacBook Pro M3 | MacBook Pro M3 Pro, M3 Max |
| Mac | M-Series 18 mo? Ultra lag | iMac M3 | Mac Mini M2, M2 Pro | Mac Studio M2 Max, Ultra Mac Pro: M2 Ultra |
| iPad | A and M 24 mo? | iPad Mini: A15 iPad: A13, A14 | iPad Air M1 | iPad Pro M2 |
| Watch | S: 12 mo | Watch SE 3/S8 | Watch 9: S9 | Ultra 2: S9 |
| AirPods | H: 24 mo | AirPods: H1 | AirPods Pro: H2 | Max: H1 (48 mo?) |
To explain what seems to be happening and the net, net is that before you buy any Apple product, you should check the MacRumors Buying Guide which tells when to buy what, and don’t buy anything that is not great or good to go. Finally, use Apple Insider Price Guide to find a good price, but if you want to understand the rhyme and reason here are some explanations
- iPhones and Watches Annual September Updates. Traditionally, Apple has updated the processors in their iPhone lineup every year, so today we have the A17 Pro chips that are in the iPhone Pros and these are the 3nm TSMC node. Then their other phones like the regular iPhone run on last year’s model (in this case the A16) and then the iPhone SE is not updated as fast probably once every 18 months and uses a version that’s two behind. If you are thinking about buying an iPhone, to maximize longevity, you do want to buy in the Fall. The positioning is that the iPhone SE is a budget phone and gets updated at a slower rate. The main line unit the iPhone 15 is probably more phone than anyone ever needs unless you are hot on photography in which case the iPhone 15 Pro Max is the one to get. Or if you really want a handheld gaming system (another niche use) the iPhone Pro/Pro Max is good. The watches are also pretty easy to understand, they are also on an annual cycle with good, better, and best pretty clear. The middle product Watch is pretty obvious as well as the value leader.
- Macs chips every 18 months (with Ultra 6 months later?) With the arrival of Apple Silicon, we saw a massive improvement in performance (on the order of 5x compared with Intel chips). The confusing thing is that they don’t announce the entire line at once and we have to take into account that the M1 had a slower rollout because it was so new. And with the pandemic, there are supply chain issues. But the best guess right now is the M1 was a “slow rollout” and the M2 had a “delayed rollout” which is why we saw the first M1 in 2020, then in the M1 Pro and Max in 2021, and with the M2, we saw the M2 in October 2022, but the M2 Max and Ultra slipped into January 2023. This makes sense from a yield point of view, the base chip is easier to make, then the Pro is a binned Max and then finally the Ultra is two Maxs glued together. But with the new M3 series, things have changed, the M3 is a real base model, the M3 Pro is more like a slightly grown-up M3 and the M3 Max is balls to the walls fast. And the M3 Ultra will presumably come out later. One impact of this is that you want to time your purchases right after the release. So the Insider Guide is really good. What is happening now is that the model changeovers are much more staggered than with iPhones so timing is everything. As an example, the surprise (see below) M3 announcement basically announced all the M3 chips (except the Ultra) so the models are very uneven. Only the MacBook Pro and the iMac got the new M3 chip, so you can expect that the MacBook Air, Mac Mini, and Studio are all going to get the M3 makeover soon. The laggard is the Mac Pro which just got the M2 Ultra, this is a hard chip to make, so I think you can expect the Mac Pro to lag quite a bit (and the uses are pretty niche right now). Performance-wise, the M3 is about 15% faster than the M2 which is in turn 15% faster than the M1, so it’s not about single-core performance. What is changing is the number of cores is rising dramatically which is good for multithreaded applications and also for gaming (as the number of GPU and CPU cores rise). The second thing is that the chips may ship in October, but the product mix changes gradually, so the best guess is the MacBook Pros (the core of their line) will get the M3 Pro/Ultra first and then one of the consumer models will (I would have thought this would be the MacBook Air, but this year it was the iMac which was late). Basically, once you know what chip you want, then you have to guess when the products migrate and hopefully post-pandemic supply chain issues, they can tighten up the releases so you don’t get a spray of different chips in different products.
- iPads a mix of A and M slipped into 2024 not synced to M or A chips. The iPads are more confusing because the product positioning is much less clear and it looks like the major products slipped from October 2023 into 2024 maybe for supply chain or development reasons. Also because this is a less important product, they tend to get the new chips offset by 12-24 months. The lower-end models use the A series in a spray of variants depending on price point and then the higher-level variants use the M series, but the operating system doesn’t really take advantage of the M performance. The speculation is that there was an aborted attempt to make the iPad the low-end computer to replace the bottom of the MacBook line, but this didn’t happen. The rough idea analysts think is that the lineup is iPad for good, iPad Air for better, and iPad Pro for best with the iPad Mini really being in a confused spot. So as an example, the iPad Pro is M2, the iPad Air is M1, the iPad Mini is A15, the iPad 10th generation is A14 and the 9th generation is A13, This means the iPad Pros are roughly 2x as fast as the lowly 9th-generation iPad, but it doesn’t matter much as not much can take advantage of it. Net, net, it’s not a great time to buy an iPad, probably 2024 will see a refresh to the M3/A17 level, but no one knows for sure. The best buy by the way seems to be the iPad Air (the middle product), but that one is really old with the M1 in it. Personally, I think buying an iPad is probably the hardest thing right now for the power user, my iPad Pro 2018 is still just fine (and it’s five years old!). The main rumor is the iPad Pro will get a really nice new OLED screen in 2024 and probably the M3 chip.
- AirPods in transition to H2 and USB C. This line is also in a strange state, the better product, the AirPods Pro is easy since it has the H2 chip, but the base AirPods are due for a refresh and the AirPods Max is in a funny position in that it hasn’t been updated and isn’t likely to be soon or at least the rumor is towards the end of 2024. These devices are also in the middle of the transition to USB C, so it makes sense to not buy a Lightening device now, just wait for the USB C versions to show up.
M3, M3 Pro, and M3 Max Announcement Surprise and Cadence changes
Yesterday, we saw the “surprise” launch (as in the pundits were predicting until October 30 that there would be no M3 Macs at all and the event was like an iPad one), but instead, we had the launch of three of them: M3, M3 Pro and M3 Max.
The other confusing thing is that the jump from Intel to M1 was huge in speed and battery life, but the M2 (which might be called the M1.5 as it is based on the same 5nm process as the M1) and M3 have been more incremental and the performance increases differ by the part of the chip involved, the top line is that there are way more cores and way more choices:
- The single-performance computer core. A way to think about this is, that the overall performance on a single-threaded application has gone up about 15% from M1 to M2 to M3. So probably about 30% now. Nice, but nothing to write home about.
- Multiple compute cores. The compute cores are typically 8 total cores with 4 performance and 4 efficiency cores in the base M1, M2, and M3, the main difference is clock rates, with the M1 running 3.2GHz/2.06GHz and the M2 at 3.49/2.4 GHz which accounts for much fo speed improvements. Net, net, if you have the previous version, it’s not like you get a huge performance improvement.
- Multiple graphics cores. This is roughly analogous to the nVidia measures of hardware which is where they talk a lot about how many GPU Shaders they have and Tensor Cores. The world of silicon means there isn’t a single measure, but various blocks on the die that have different functions. For instance, GPU Shaders for games vs Tensor Cores doing AI vs Ray Tracing cores. Similarly, Apple Silicon has graphics cores, Neural Engines, and so forth.
Like Intel, Apple has a large gradation of cores and memory that you can choose from, the most important variable if you are doing machine learning or video editing though is the amount of RAM to fit models in, otherwise, a base 16-GB machine is really pretty fine for most ordinary uses. But please don’t get the 8GB if you are getting an M3 Pro. That’s fine for a low-end MacBook Air that is just for browsing and videos.
Understanding M3, M3 Pro and M3 Max performance
This could actually fill up an entire blog post and in fact Tom’s Guide has done that
One important thing to know the M3 Pro is detuned from the M2 Pro (it is less like an M3 Max minus and more like an in-the-middle product now. What that means is that is less of a good buy than in the old days). For instance, its memory bandwidth is 150GBps vs 200GBps in the M1 and M2. It has one less GPU core (18 vs 19 in the M2 Pro). As an aside these figures ultimately have to do with how many memory banks you have and how many PCIe channels there are out of the processor. As an example, the M3 Pro is at 36GB because it has three banks of 12GB RAM each. And it looks like they have a bunch of different chip configurations and also binning, so good luck figuring out the memory vs processor grid. It’s pretty sparse.
Another example is that the M3 Pro has 6 performance cores (and 6 efficiency), but the M2 Pro has 8 performances (and 4 efficiency). Also, they did some more “binning” so the M3 Max has two versions. A 14-core CPU/30-core GPU has 300GBps of bandwidth, but the M2 Max has fewer cores 12-core CPU, but has the same 400GBps bandwidth as its bigger brother.
Net, net these are small changes and it is not clear what real effect all this binning of different versions has on performance (and I’m sure performance will vary by application). But the first benchmarks are coming out now. For instance, nanoreview.net has some detailed benchmarks as an example, but this compares:
| Feature | M3 Pro (12-core) | M3 Max (20 core) | Improvement |
| Performance Cores | 6 @ 3.7GHz | 12 @ 3.7GHz | 2x |
| Efficiency Cores | 6 @ 2.4 Ghz | 4 @ 2.4GHz | 1/3 less |
| RAM Multiplier | 37x | Same | Same |
| Caches | 192K/core, 32MB shared | Same | Same |
| Transistors | 37B | 92B | 2.5x |
| GPU | 19-core | 38-core | 2x |
| Max Memory | 36GB | 128GB | 3.5x |
| Memory Channels | 4 | 8 | 2x |
| Bandwidth | 153.6GBps | 409.6GBps | 2.7x |
But as usual, when you do real-world performance it is not as dramatic, looking at this you would expect say 33-40% more performance with real-world application for multithreaded. Since the clock rates are basically the same for single-core metrics, the only real difference is the memory bandwidth which is about a 5% improvement. Cinebench is a bunch of synthetic graphics applications while Geekbench is more general it has file compression, C compilation, browser rendering, PDF and Text as well as a bunch of photo things like background blur and ray tracing.
| Benchmark | Core | M3 Pro (12-core) | M3 Max (20-core) | Improvement |
| Cinebench R23 | 1 | 1878 | 1963 | +5% |
| Geekbench v6 | 1 | 3083 | 3227 | +5% |
| Cinebench R23 | Multi | 17534 | 23272 | +33% |
| Geekbench v6 | Multi | 14982 | 21167 | +41% |
Now 40% faster to do things like video rendering is actually quite a big deal, so it depends on how much you do these monster tasks, but most mortals aren’t going to need an M3 Max 🙂 Note that GEekbench relies on crowdsourced data, so it can be both more and less accurate compared with a single lab sample, YMMV, but a quick look at Mac 15,9 which is the new M3 Max and the overall Mac Benchmarks provides some perspective where we scale the measures by the M3 Pro’s 3083 and 14982 numbers for single and multicore. The main thing that you can see is that for single core what a huge jump it was from Intel and the performance gains have been relatively modest compared with the M2. So to put it in real-world terms, compared with the M2 of any version it is about 1.1-1.2x faster, so a 30-minute job would take 25-27 minutes as an example. Then compared with Intel is is about 1.3x faster. Note that even the fancy Ultra and Max chips are not much faster because they have more cores, not more clock speed.
Finally compared with the Intel chips there are real differences at 2.2-2.4x, so a 30-minute job on Intel would take 12-13 minutes and of course use way less power. And the M3 Ultra is not much faster
| Single Core Geekbench | Processor (CPU@PClock, GPU) | Score | M3 Pro |
| MacBook Pro 15″ (2023) | M3 Ultra (16@3.7, 38) | 3227 | 0.95x |
| MacBook Pro 15″ (2023) | M3 Pro (12@3.7, 19) | 3083 | 1x |
| Mac Studio (2023) | M2 Max (12@3.7Ghz, 30) | 2803 | 1.1x |
| Mac Studio (2023) | M2 Ultra (24@3.7GHz, 60) | 2768 | 1.1x |
| MacBook Pro 16″ (2023) | M2 Max (12@3.7GHz, 30) | 2736 | 1.1x |
| Mac mini (2023) | M2 Pro (12@3.5, 19) | 2649 | 1.2x |
| MacBook Pro 14″ (2023) | M2 Pro (10@3.5, 16) | 2638 | 1.2x |
| Mac mini (2023) | M2 (8@3.5, 10) | 2631 | 1.2x |
| Macbook Air 15″ (2023) | M2 (8@3.5, 10) | 2590 | 1.2x |
| Mac Studio | M1 Max (10@3.2) | 2402 | 1.3x |
| MacBook Pro 14″ (2021) | M1 Pro (10@3.2) | 2375 | 1.3x |
| Mac Mini (2020) | M1 (8@3.2) | 2350 | 1.3x |
| MacBook Pro 15″ (2019) | Core i9-9980HK (8@2.4) | 1393 | 2.2x |
| MacBook Pro 15″ (2018) | Core i7-8850H (6@2.6) | 1264 | 2.4x |
| MacBook Pro 15″ (2017) | Core i7-7820HQ (4@3.1) | 1178 | 2.6x |
Then for Multicore things are spread more because there are more cores in later designs and the memory bandwidth is going to make more of a difference as are more GPUs assuming that with single core the benchmarks are not saturating memory or GPU and this is what happens, the differences do expand. As an example, the performance advantage of the M3 Ultra is much faster for multithreaded, a 30-minute job on a M3 Pro takes 21 minutes. As an aside this is nice but not a huge difference unless you are doing hour jobs.
And also the performance of the M2, M2 Pro, and M2 Ultra spread quite a bit. As an example, the M2 Max is actually about the same speed as the M3 Pro while the M2 Ultra keeps up with the M3 Max. And you can see that the M3 Pro is now 10% better than the M2 Pro (vs 5% in single-threaded), so that’s why they are doing all their comparisons against the M1s, the M2 and M3 difference is really not that much even in multithreaded use.
Finally, the difference between the M3 Pro and regular M2 is pronounced at 1.5x faster. However, even with the now-old M1 line, it’s not really that much. The M1 Max and M1 Pro are about 1.2x slower. But the differences with Intel are now huge in the 2.4x-4x range. That means a 30-minute job on M3 Pro would take 1-2 hours on an Intel machine.
| Multicore Geekbench | Processor (CPU@PClock, GPU) | Score | M3 Pro |
| MacBook 15″ (2023) | M3 Max (16@3.7, 38) | 21167 | 0.7x |
| MacBook 15″ (2023) | M3 Pro (12@3.7, 19) | 14982 | 1.0x |
| Mac Studio (2023) | M2 Max (12@3.7Ghz, 30) | 14805 | 1.0x |
| Mac Studio (2023) | M2 Ultra (24@3.7GHz, 60) | 21318 | 0.7x |
| MacBook Pro 16″ (2023) | M2 Max (12@3.7GHz, 30) | 14497 | 1.0 |
| Mac mini (2023) | M2 Pro (12@3.5, 19) | 14254 | 1.1x |
| MacBook Pro 14″ (2023) | M2 Pro (10@3.5, 16) | 14226 | 1.1x |
| Mac mini (2023) | M2 (8@3.5, 10) | 9743 | 1.5x |
| Macbook Air 15″ (2023) | M2 (8@3.5, 10) | 9743 | 1.5x |
| Mac Studio | M1 Max (10@3.2) | 12433 | 1.2x |
| MacBook Pro 14″ (2021) | M1 Pro (10@3.2) | 12202 | 1.2x |
| Mac Mini (2020) | M1 (8@3.2) | 8390 | 1.8x |
| MacBook Pro 15″ (2019) | Core i9-9980HK (8@2.4) | 6406 | 2.4x |
| MacBook Pro 15″ (2018) | Core i7-8850H (6@2.6) | 4896 | 3.0x |
| MacBook Pro 15″ (2017) | Core i7-7820HQ (4@3.1) | 3730 | 4.0x |
Conclusions: MacBook Air for mostMove to Pro M3 if you have Intel or are doing ML
Well, the main conclusion is that if you have an M1 or M2 MacBook that you like, you are good to go. And if are on Intel MacBooks right now, a good thing to wait for will be the M3 MacBook Airs, those are going to be a nice upgrade. Then for power users, if you have M2, you are not really far behind, and even with M1, the performance differences are not huge. This is actually good news in that the longevity of the MacBooks seems really high right now.
If you have an M1 or M2, its kind of on the bubble, the performance differences aren’t great, the biggest reason I’ve seen that people are upgrading is the need to do machine learning and AI development where they want to run models locally (way faster than doing it in the cloud).
