Apple M3, M3 Pro, M3 Max

[GR Baumann]

M3 Max pushes the transistor count up to 92 billion and takes pro performance to the next level. The 40-core GPU is up to 50 percent faster than M1 Max, and support for up to 128GB of unified memory allows AI developers to work with even larger transformer models with billions of parameters. The 16-core CPU features 12 performance cores and four efficiency cores, achieving astonishing performance that’s up to 80 percent faster than M1 Max.

So here they are, the first three nanometer chips. MacBooks and iMacs only for the moment. The iMac with 8GB Ram and 512 GB SSD is 2.059,00 € in Germany, and can have a maximum of 24GB & 2TB.



Overall prices increased as well.

The flagship 16" MacBook Pro M3 Max
16-Core CPU, 40‑Core GPU
48GB & 1TB
4.849,00 €

maxed out spec
128GB & 8TB
your eyes bleed at the 8.529,00 € price tag.

The cheapest 14 inch M3
8‑Core CPU, 10‑Core GPU
8GB & 512 GB SSD
1.999,00 €.

M3 Pro,
11-Core CPU, 14‑Core GPU,
18 GB & 512 GB SSD
2.499,00 €.


Now here is the crux between M3 Pro and M3 Max, the Pro goes to 36 GB only.

96 GB is available with the M3 Max with 30-core GPU. 48GB, 64GB or 128GB are available with the M3 Max with 40-core GPU

Check that out,
MacBook Pro 14":
14‑Core CPU, 30‑Core GPU, 96GB & 1TB 4.919,00 € (16" 5.169,00 €)
vs.
MacBookPro 16":
16-Core CPU, 40‑Core GPU, 48GB & 1TB 4.849,00 €

I know what I would do if I had 5 grand to dish out on such a beast.

https://www.apple.com/newsroom/2023/10/ ... -m3-chips/

[Guy Rowland]

A cautious sense of relief here from an M2 owner. They've gone for a split between performance and efficiency cores which IIRC has not been advantageous for audio use in the case of Intel's range (EDIT - they already had this, but with some versions there's now more efficiency vs performance cores). Their boasted speed improvements are actually not that wild in most cases - sometimes as low as 30% - and this is comparing with the M1, not the M2. So if that's the audio-use benefit as I suspect, then it's more or less a similar modest improvement from M2 to M3 as M1 was to M2.

But a lot of assumptions there on my part. It's great for AI-related tasks, I read. Until it's in the real world running a DAW it's hard to know if this is just an incremental improvement, rather than the touted 3nm quantum leap that we saw in the jump from intel to silicon. Also battery life improvement seems similarly modest - it's huge compared to Intel, but only an hour or so compared to M2 IIRC.

The configurations are weird - the amount of RAM you can have varies from specific model to specific model, not just the difference between M3, M3 Max and M3 Pro (and what's all this business about 36GB?!). As for cost, it feels UK-wise it's a bump but not a massive one - this stuff always was incredibly expensive.

Perhaps from our point of view, one of the main things that struck me is that in 2023 we're still having to be very cautious on resource use. There are restrictive finite limits on RAM and storage space in the mobile arena. What you get is very expensive, and there's big constraints on what you can get. So for developers, my plea to think about install options like being able to install a 1-mic version of a 15-mic library is very much pertinent. And all the tricks we have using purged samples, disabled tracks etc are still going to be very useful for a long time to come.

I think it's gonna be a while til we really know what we're dealing with regarding the M3 series. One thing's for sure, speccing a new macbook system just got a lot more complicated.

[GR Baumann]

Affirmative, I see it the same way with my MacMini M2, that being said, I wonder what's in the pipe for the mini. The iMac is strange isn't it, very very limited. I think we were lucky to have jumped on the M2 train. I was considering to sell the M2 and eventually get a M3 Pro Mini, but seeing now it likely would be limited to 36 GB, nah, makes no sense for what, 30% more CPU power at a cost of another 1.5k minimum. Nope, I think the M2 will serve me fine the next few years for my hobby therapy.

[Guy Rowland]

Affirmative, I see it the same way with my MacMini M2, that being said, I wonder what's in the pipe for the mini. The iMac is strange isn't it, very very limited. I think we were lucky to have jumped on the M2 train. I was considering to sell the M2 and eventually get a M3 Pro Mini, but seeing now it likely would be limited to 36 GB, nah, makes no sense for what, 30% more CPU power at a cost of another 1.5k minimum. Nope, I think the M2 will serve me fine the next few years for my hobby therapy.

...and I'm not at all sure it would be a 30% increase over your M2 for audio work. So I agree, very much NOT an I-must-have-this moment, but for new buyers or those still on Intels etc it will of course be incredible.

I hate the clickbait still image and text on this video, but it actually seemed a good analysis on all the various pros and cons, what we do and don't know:

[GR Baumann]

The interesting thing will be unified memory, I can imagine Logic and Kontakt to benefit from that, to what percentage compared to M2, time will tell.

[Lawrence]

WTH is “unified memory” anyway? I understand that RAM isn’t exactly RAM anymore, but I don’t get how it works.

[GR Baumann]

It is supposed to be even more efficient in the M3 Larry, memory resource available for both the CPU and GPU.

[mcalis]

WTH is “unified memory” anyway? I understand that RAM isn’t exactly RAM anymore, but I don’t get how it works.

It refers to the CPU and GPU using the same (physical) memory. In plain English: instead of having two separate cupboards, one where the CPU is allowed to store things and one where the GPU is allowed to store things, there is now one big cupboard that both the CPU and GPU use.

The M1/M2/M3 have the CPU, GPU and all the memory all on a single chip, a single physical component instead of distinct, separate components that are located on different parts of the motherboard.

Mind-blowing little fact: the physical distance, literally the centimeters or millimeters where the memory is located on the board has an impact on the speed performance of the memory. A big part of why M1 and M2 are so fast is because the memory is located right on top of the CPU, so the physical distance from the CPU to the memory is shorter and so the electrical signals need to travel less far, hence the entire thing becomes faster. The obvious tradeoff is that if there is any kind of failure anywhere in the chip, the whole system is toast. In a regular desktop PC, the CPU, GPU and RAM are all physically separate components so a failure of one means only that one component would need replacing. With an M1/M2/M3, if the chip fails, your entire computer is dead.

[Guy Rowland]

Thanks Matthias, that's a super-helpful explanation.

[Luciano Storti]

WTH is “unified memory” anyway? I understand that RAM isn’t exactly RAM anymore, but I don’t get how it works.

It refers to the CPU and GPU using the same (physical) memory. In plain English: instead of having two separate cupboards, one where the CPU is allowed to store things and one where the GPU is allowed to store things, there is now one big cupboard that both the CPU and GPU use.

The M1/M2/M3 have the CPU, GPU and all the memory all on a single chip, a single physical component instead of distinct, separate components that are located on different parts of the motherboard.

Mind-blowing little fact: the physical distance, literally the centimeters or millimeters where the memory is located on the board has an impact on the speed performance of the memory. A big part of why M1 and M2 are so fast is because the memory is located right on top of the CPU, so the physical distance from the CPU to the memory is shorter and so the electrical signals need to travel less far, hence the entire thing becomes faster. The obvious tradeoff is that if there is any kind of failure anywhere in the chip, the whole system is toast. In a regular desktop PC, the CPU, GPU and RAM are all physically separate components so a failure of one means only that one component would need replacing. With an M1/M2/M3, if the chip fails, your entire computer is dead.

Thanks for the explanation, Matthias! Does that mean that in an SoC like the Mx there is no longer a need for a cache? Since the distance is now minuscule is there even a need to store smaller bits of data nearby?

[GR Baumann]

>>With an M1/M2/M3, if the chip fails, your entire computer is dead.<<

In deed, which is why even for my MacMini I opted for the Apple care plan.

Luke, there is a lot of mystery surrounding cache management in the M3 chips afaik.

This is from Apple:

“Dynamic Caching, unlike traditional GPUs, allocates the use of local memory in hardware in real time. With Dynamic Caching, only the exact amount of memory needed is used for each task. This is an industry first, transparent to developers, and the cornerstone of the new GPU architecture. It dramatically increases the average utilization of the GPU, which significantly increases performance for the most demanding pro apps and games.”

Further: https://www.tomshardware.com/software/m ... -for-years

[Luciano Storti]

Luke, there is a lot of mystery surrounding cache management in the M3 chips afaik.

This is from Apple:

“Dynamic Caching, unlike traditional GPUs, allocates the use of local memory in hardware in real time. With Dynamic Caching, only the exact amount of memory needed is used for each task. This is an industry first, transparent to developers, and the cornerstone of the new GPU architecture. It dramatically increases the average utilization of the GPU, which significantly increases performance for the most demanding pro apps and games.”

Thanks Georg. Looks like this specifically discusses the GPU architecture. But I suppose, then, it would have to be relevant since it's a SoC. Was wondering specifically about CPU L1, L2, L3 caches.

[GR Baumann]

afaik for the M3 unknown to date, not?

I saw a chart comparison six days ago where only ? were set for cache in the M3 series p- and e- cores.

Interesting though, Apple compared the M3 mostly to the M1, which leaves me to believe that the M3's goal was to reduce production cost mainly, the design is a hint here too, and the performance difference to the M2 series, raytracing abilities aside, are probably not as big as one would have expected. The CPU performance jump from M1 to M2 is probably greater than from M2 to M3, but it is early days.

[mickeyl]

Thanks for the explanation, Matthias! Does that mean that in an SoC like the Mx there is no longer a need for a cache? Since the distance is now minuscule is there even a need to store smaller bits of data nearby?

No. L1/L2/L3-Caches are still a necessity, but the impact of a cache miss is way less than on traditional architectures.

[Luciano Storti]

Aha, thanks.

[GR Baumann]

Saw a comparison with Intel, and on the grounds of CPU per watt, with the latest intel with 9+ flippin Ghz (10nm) sucking somewhat around 600 watt, which was 10x more power used than the fastest M3 (3nm ), who ran on batteries on top, which the intel can not without loosing a lot of CPU performance, and still the M3 smoked that i9-14900KF in certain tasks.

Why are monstrosities like an i9-14900KF still brought to market and more crucially, why people purchase such? Oh wait, there is that thing with fossil fuel powered cars etc as well... yeah, of course, fucking humans.

[mickeyl]

Well, the achilles heel of the shared memory architecture is obviously the scalability with regards to RAM size. For applications where you need terabytes of RAM, there is just no alternative (yet) to Intel, AMD, and the like.

When Apple moved its Mac Pro from an Intel-based design to their own silicon, they transitioned from a maximum RAM size of 1,5 Terabyte (!) to 196 Megabyte. And while this is still impressive per se, for special applications, it just doesn't cut it.