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.
Macs to Get AI-Focused M4 Chips Starting in Late 2024
- Thread starter MacRumors
- Start date
- Sort by reaction score
I’m a professional and I never want to use a separate webcam for video calls again unless filming something for broadcast or exhibition. And I’m glad Apple used the notch because the 1080p quality of the new cam is pretty good. It you didn’t have a notch you’d have a whole bar along the top of the display which would increase the size/weight…
They better release the M4 Max with 256 GB support, I might upgrade from my M1 max MBP16 a year early, if the RAM goes up, and also if it can match 50-70% of Nvidia XX90 GPU.
But then you’d have less screen real estate. They do black out the status bar when you’re browsing the web.
You are saying that they are behind based on Copilot? Hahaha. Copilot is mainly annoying and has caused advertising even in places like the developer tools for Edge. It is just great to be reviewing the communication from the browser to the server and have to read the ads that are interjected in the middle of it. AI will be great once we get past the spin cycle for it.
of interest to the NPU discussion
see https://www.notebookcheck.net/Apple...-the-A17-Pro-A16-Bionic-designs.764280.0.html
One aspect of the provenance of the M3 design that is crystal clear is that it shares the same, or very similar, GPU architecture to the A17 Pro. Apple calls the A17 Pro graphics “the biggest GPU redesign in Apple’s history,” while it says the M3 graphics “represents the biggest leap forward graphics architecture ever for Apple silicon.” With the A17 Pro launched on 13 September and the M3 chips launched on 31 October, it is safe to assume the architecture is going to be very similar, especially given the brand new hardware accelerated ray tracing feature is common to the two. The only exception may apply to the “Dynamic Caching” feature in the M3 GPU that wasn’t called out as a feature of the A17 Pro GPU. Otherwise, the A17 Pro and M3 GPUs share much more in common than do the A16 Bionic's GPU and the M3 GPU.
Things change, however, when we look at the M3 series NPU (Neural Processing Unit). As in previous M series chips, the M3 “Neural Engine” (as Apple calls it), has 16 cores and produces 18 TOPS (trillion operations per second). This is almost 2x slower than the NPU in Apple’s latest smartphone chip, which is somewhat unexpected. The NPU in the Apple M1 is good for 11 TOPS (the same as the NPU in the A14 Bionic), while the NPU in the Apple M2 was 15.8 TOPs (the same as the NPU in the A15 Bionic). As you might have guessed, the NPU in the M3 is much closer to the NPU in the A16 Bionic’s 17 TOPS, indicating that this is the basis for the M3 series NPUs. The 1 TOP differential can be explained by the performance gained through the switch to the 3nm node used for the M3 against the 5nm (N4P) node used for the A16 Bionic. Or it could simply be the result of a clock boost.
As for the CPU architecture, that is a little more difficult to discern, but there is a very good case for it to be made that it is more likely to be in line with the A16 Bionic. While the shared NPUs is one clue that the CPU architecture is also the same, the next-generation GPU architecture that we have now seen debut on the Apple A17 Pro and M3 series was actually originally intended for the A16 Bionic. An exclusive report [sub.req] from The Information in December of 2022 said that Apple’s engineers were “too ambitious” with this ray tracing-capable GPU. Prototype A16 Bionic chips running the new GPU architecture suffered from overheating and drew too much power. This “unprecedented” blunder came amidst an exodus of top chip talent from Apple, including former A series lead architect Gerard Williams who left to found Nuvia and is now SVP of Engineering with Qualcomm following its acquisition of Nuvia. Williams was recently on stage at Qualcomm’s Snapdragon Summit to launch the impressive M2 Max-beating Snapdragon X Elite chip.
What this is suggestive of is that the M3 series chips were probably intended to be entirely based on the A16 Bionic, including its proposed next-generation GPU. Apple pivoted with the A16 Bionic by improving graphics memory bandwidth to boost performance over the A15 Bionic, but delayed the implementation of the new GPU architecture until the A17 Pro. Apple will have undoubtedly had all the technical documentation and schematics laid out for the M3 with A16 architecture - including the new GPU architecture. However, where they had to rush a fix for the A16 GPU, they would have still had time to get the GPU sorted out as planned for the M3 series, while also making sure it was good to go for the A17 Pro.
Given the M3's NPU connection with the A16 Bionic - and the fact that what was supposed to be the A16 Bionic GPU ended up in the M3 - it seems highly likely that both the A16 Bionic CPU and NPU architecture also form the basis of the M3. Thus, the Apple M3 series looks to be a hybrid of the A16 Bionic and A17 Pro, but an accidental hybrid. It received the CPU, NPU and GPU cores always intended for the chip - just by what looks to be a more circuitous route.
Exactly. The NPU needs to get the same treatment the GPU does with having more cores per model level, at least. There are going to be a lot of annoyed M3 Mac Studio / Mac Pro buyers (assuming those products get updated with M3) if they care about this type of performance because laptops coming in 6-8 months are going to wreck them on that front.
Apple did a great job with fast large pools of unified memory aiding ML training and the M1 was somewhat competitive at the time to the point that they highlighted it as the best laptop for ML/AI development, and it's been notably (and rightfully) absent in their marketing since then because it isn't competitive unless your model can't fit into memory. It's a joke that the iPhone is faster than the M3 at some operations, but the M2 and M3 are weird products on a lot of levels. The DMP needs a slight revision to close the security hole with regard to memory addressing, plus they could really use things like Thunderbolt 4 and AV1 encoding and are behind the curve there.
The M4 will be pretty great assuming these things are addressed, and I think they will be.
Last edited:
Unlike sensor size (ie full frame, APS-C, 1", etc) resolution has nothing to do with the size of the camera. Regardless of how many pixels you cram into a specific sensor size the required image circle, hence lens size, remains the same. F-stop and focal length do have an impact on required size.
Transformers, the neural network architecture that underpins LLMs do have contextual understanding. This from the multiheaded attention module which calculates dot products for each token, for every other token from a dataset that has been tokenized (this can be strings, images, etc).
now if you want to argue that this particular method of contextual understanding is flawed, we can, but even older models before transformers (RNN, seq2seq) do feature contextual understanding. i suspect you mean that this "contextual understanding" is based on patterns in a dataset, rather than actual comprehension; and I would agree with you on that, but these models are much more than "glorified" statistics.
Chips designed for AI, like neural processing units (NPUs) or neural engines, are specialized hardware optimized for running AI algorithms efficiently. They accelerate tasks like deep learning, which powers many AI applications today. These chips can perform calculations in parallel, which is crucial for handling the massive amounts of data involved in AI tasks.
Neural engines specifically are designed to simulate the behavior of neurons in the brain, enabling faster and more efficient processing of neural networks. Enhanced AI refers to the improvement in AI capabilities achieved through advancements in hardware, software, and algorithms working together. It’s not just about processing requests but also about learning from data and making predictions or decisions. So, while supercomputers are still used for some AI tasks, specialized chips bring AI capabilities to smaller devices like smartphones, making AI more accessible.
M4 chips will not be significantly larger or have significantly more transistors. So if the M4 NPU will be a significant enhancement, what is going to be sacrificed? Fewer GPU?
Or perhaps the significance of the update for M4 is more just fancy marketing.
Or perhaps the significance of the update for M4 is more just fancy marketing.
We are up against people’s unrealistic expectations of how much improvement each generation of chip can provide and a tendency to think that a 15-20% improvement is lame. M1 was a huge jump because of where the pre-AS Apple laptops were with their Intel chips. Once we jumped up to the new plateau, each step up from that was never going to be such a big increase.
So you’re the glass half empty type?
The notch allowed Apple to provide more screen space. You can turn that upside down and see it as a reduction but it just ain’t so.
Relax, there have always been computer and processor updates, typically annual and sometimes more often. In most cases, the year over year difference isn’t huge. Like on the phones, you need to wait 2,3, or 4 years and then an upgrade really does feel like an upgrade. The annual upgrades are about keeping up with competition and providing upgrades for people who still have older machines.
It only feels a little overwhelming if you spend a lot of time on a rumor site like this where it is a constant stream of new new new. For normal people who might be thinking of a new computer, none of the rumors mean much. They just go buy what they need when they need it.
Transistors are even bigger with N3E. So they take the NPU cores from A17 pro which gives a 100% speed boost including the higher clock speed vs. M3. They should reach ca. 80 TOPs with M4.
Nothing really I just felt weird buying An M2 when the M3 was already out. I think Apple should rethink the way they release these things now if they wanted to skip M3 and throw an M4 in the next max studio I think that would be great.
So if this is true, Apple is going from 4 processor tiers (e.g. Mx, MxPro, MxMax, MxUltra) to just 3. So the new mid-tier will be taking the place of both MxPro and MxMax?
While there is some overlap with the current 4 tiers (e.g.; between a loaded MxPro and lowest level MxMax). And there is some logic for greater delineation between tiers, this is a big departure from how Apple has been marketing its hardware (just look at the ranges of iPads)
Last edited:
The
The Middle chip powers the higher-end macbook pros.
The High end chip goes into the Desktop (Mac Studio and Mac Pro).
A couple ways this could go.
1. The MxPro and MxMax are somehow one chip, and the high end M4 is the M4Ultra. This seems wasteful and unlikely.
2. There's an M4Ultra chip that is on a later path. Like there's an M3,M3Pro,M3Max today and no M3Ultra.
The M2Ultra was 2xM2Max with an interposer, so perhaps M3Ultra is coming soon too (and it would be a monster).
In this universe, the article was a little off and we still have the mid and high end chips powering the high-end Macbook Pros. Then the Desktops will run M4Max and 2xM4Max (M4Ultra).
Now there's still the matter of the reported maximum memory of 512GB for the desktop machines. Will the M3Max have its maximum memory moved up to 256GB? That would be exceptional for a laptop, yielding an M4Ultra (2XM4Max) with 512GB. Or has Apple figured out a way to enable a 4XM4Max configuration?
Other ideas?
The Bloomberg article said the low end chip goes into the Macbook Air, low-end Macbook Pro, and the Mac Mini.
The Middle chip powers the higher-end macbook pros.
The High end chip goes into the Desktop (Mac Studio and Mac Pro).
A couple ways this could go.
1. The MxPro and MxMax are somehow one chip, and the high end M4 is the M4Ultra. This seems wasteful and unlikely.
2. There's an M4Ultra chip that is on a later path. Like there's an M3,M3Pro,M3Max today and no M3Ultra.
The M2Ultra was 2xM2Max with an interposer, so perhaps M3Ultra is coming soon too (and it would be a monster).
In this universe, the article was a little off and we still have the mid and high end chips powering the high-end Macbook Pros. Then the Desktops will run M4Max and 2xM4Max (M4Ultra).
Now there's still the matter of the reported maximum memory of 512GB for the desktop machines. Will the M3Max have its maximum memory moved up to 256GB? That would be exceptional for a laptop, yielding an M4Ultra (2XM4Max) with 512GB. Or has Apple figured out a way to enable a 4XM4Max configuration?
Other ideas?
Last edited:
1) Looking back at the source and questioning Gurman”s accuracy?
2) Expanding the idea of chiplets and interconnects, where a tier 2.5 chip is based on interconnection between multiple mid or low tier chips.
3) Binning of higher level chips (or lower clock rate)
well, that’s settled then, my next upgrade from my current M1 Mini is going to be the M4 Studio. If I can get my hands on one, I bet this upgrade cycle is going to be huge.
I still hope Apple is also working on efficient AI dedicated server chips, they sure need them.
I still hope Apple is also working on efficient AI dedicated server chips, they sure need them.