Maynard, any chance you are getting one of the new machines? I would be very interested in your analysis of the new larger CPU clusters.
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.
M3 Chip Generation - Discussion Megathread
- Thread starter scottrichardson
- Start date
- Sort by reaction score
Maynard, any chance you are getting one of the new machines? I would be very interested in your analysis of the new larger CPU clusters.
No, don't get carried away
If I have "documented" anything, it's that there's a set of patents that show Apple has been thinking very seriously about how they could grow the design, both at the physical level (how to connect the pieces to each other and the RAM to the SoCs) and at the logical level (how to design things like cache, interrupt and power control to scale over multiple SoCs).
But it's always risky to try attaching a timeline to these things. What seems like the obvious "right time" from the outside may be unaware of some killer problem on the inside that means that "we'll just have to skip scaling up this year, and fix the problem with the M4".
The real constraint Intel Macs impose is on the GPU side.
It's "fairly easy" for Apple to recompile changes in the OS and UI for Intel and users just get what they get, slow though it might be. But forcing Metal to continue to maintain functionality that tries to match Apple Silicon GPUs to AMD is a big problem that's only going to get worse.
Going forward I imagine Apple want to see a future of things like "single source file" GPU code, where I write one Swift function that, by adding a keyword (eg "kernel") gets compiled to run either on GPU or multiple CPUs or even ANE (kinda the same level of abstraction that you had with OpenCL, or that you have with Apple's neural network API's).
But you can't get to that point if your code has to include a whole lot of boilerplate which only exists to paper over the differences between AMD and Apple Silicon.
Probably not soon. I imagine dougall will attack one soon, but I'm supposed to be out of the benchmark business and working on other things. We'll see; this stuff is so fascinating it has a way of dragging you back in!
You're assuming an Apple keynote exists to tell you things you find interesting. Wrong wrong wrong!
An Apple keynote is a marketing device. It exists to sell to people who want to buy; and the most likely pool of buyers remains people with an existing Intel Mac. THAT is why Apple essentially keeps saying "this machine is MUCH nicer than your existing mac – go to an Apple store and try it out, or ask your friend who already have one".
That was more of a chopped design. Higher level of floorplan layout reuse layered on top of common subcomponent clusters/sections. However, as each die design sells more it gets easier to amortize the individual unique design overhead over more SoCs sold. The M3 Pro eventually getting to a Mini Pro that sells in significant numbers should help pay for making the M3 Pro more unique. ( plain Mn obviously just dominates in number of units sold. It is spread over multiple products and product lines ( iPad and Mac ). ). If the M3 Pro goes into another Mac product or iPad variant it just gets easier to spread the incremental costs out over more units.
The Mini Pro might be enough, but I wouldn't be surprised if there was something else coming over next year. ( and no , it is not the 24" iMac as several threads have tried to push. )
Additionally, I think the M3 Pro is a better match to the MBP 14" chassis now and the M3 Max likely is better matched to the MBP 16" (and definitely better matched to the Mac Studio in the more upscale configurations.) .
The suggestion that this is an AMX unit is interesting, but unclear to me. Maybe someone with better geometric skills can make a convincing case.
Issues that I think need to be considered:
(a) comparison with AMX unit on M1/M2 generation
(b) comparison with AMX unit on M3 vs M3 Pro
(c) does Apple give us one AMX unit per cluster (so one P-level AMX unit for Pro) or one per subcluster (ie two AMX units per Pro)? I could see things going either way.
AMX from the start was able to maintain separate state for all four cluster client cores, so that four cores could simultaneously execute AMX code. (This wouldn't give you 4x speed up, but would give you a small speedup as some degree of data movement could happen in parallel with a full AMX execution per cycle).
Over time Apple has done various things to both speed up one core's use of AMX (eg pack more instructions into a single transaction from the core to the AMX unit) and to make AMX more generalized (eg able to execute AVX-512-like vector operations).
IF Apple sees the main use case as matrix multiply, then there's a case for given each P sub-cluster it's own AMX unit, so Pro retains two P-level AMX units. But if Apple sees the main use case vector operations, AND if the unit is able to execute something like 4 or 8 vector operations per cycle (the baseline hardware is there, but in prior designs scheduling only allowed for, I think, two vector operations per cycle) then Apple might feel this improvement in AMX throughput is enough to allow one AMX to cover six cores?
Point is the M3 and M3 Pro "supposed AMX units" look rather different to me. And not in a way that makes sense, like the M3 Pro one is two copies of the M3 one. So I'm not convinced they're AMX, though I don't reject the idea out of hand. (Whereas they absolutely looks nothing like an E cluster to me, for so many different reasons, starting with why the Max gets two of them!)
There are other interesting things to investigate here. Right now, as I said, an AMX unit carries four copies of state, one for each client core. That's a horrible waste of register storage if only one client is using AMX (and gets worse if there are SIX copies of state...). There is a simple way to get some additional value, which may be implemented in the A16, where "speculative" AMX loads are "pre-executed" early in the pipeline to load data into some of those unused registers, then when it's time to really execute the load the data is already present and the register just has to be renamed.
This is a clever idea, and a first step towards a more "dynamic" (ie OoO) design; and presumably Apple will slowly move in that direction. The more they move that way, the more all these "excess" registers are no longer just wasted space on the die...
I'd be very curious to see AMX performance numbers for M3 Pro, both for matrix multiply and for large vector operations like FFT. Maybe we will be pleasantly surprised?
Remember every stupid lawsuit about SSD or PCIe performance is one more reason they don't talk about this stuff.
There's no upside in saying that SSDs are faster if the press story for the next month is "SCANDAL! SSDs in cheap Macs are slower than in expensive Macs!" or "Experts say SSDs in Macs, even though they start fast, will slow down over five years"...
Remember all those claims that M1 SSDs were going to die soon because of supposed massive amounts of swapping? have we seen ANY M1 SSDs die in this way? But that doesn't stop the same people who squawked loudly then from squawking just as loudly today.
If the M2 Pro floorplan and basics forked off the Max, then maybe there is a chance the desktop mulitple die SoC make a bit of divergence from the Max. In other words, don't put the UltraFusion connector on the M3 Max die at all. That the die shot delivered isn't photoshopped smaller .
I'm not saying Apple dumped the memory controllers and iGPU and building a Threadripper 'killer'. Just that floorplan is different. I/O (Thunderbolt , etc ) attached why???? Somebody needs 16 Thunderbolt controllers? Probably not.
36P cores and 384 (with no ECC ) is still rather in the fanatical range. 512GB with no ECC doesn't make much sense.
P.S. would depend upon how well the Mac Pro and Max versions of MBP did though. The same
Max die' shared between MBP and Mac Studio would cut the numbers for those Ultra(and more) numbers pretty low to amortize costs over.
Last edited:
Looking for advice and confused by the discourse around the chips:
If I’m coming from an Intel 2018 base model, should I care about the difference between 12CPU/18GPU M3 Pro and the M3 Max for video editing?
Like if I’m not sure I need it and don’t do 3D work, am I probably a good fit for the upper M3 Pro chip? I use DaVinci Resolve and some effects, but not crazy or 3D.
If I’m coming from an Intel 2018 base model, should I care about the difference between 12CPU/18GPU M3 Pro and the M3 Max for video editing?
Like if I’m not sure I need it and don’t do 3D work, am I probably a good fit for the upper M3 Pro chip? I use DaVinci Resolve and some effects, but not crazy or 3D.
I called Apple today and they went ahead and extended my return window of my 14" MacBook Pro to get the new one. But now I'm feeling cold feet due to the differences between the models lol...
The new Max design raises some interesting questions.
Suppose the following statements are both true:
- building and testing an EUV mask set is extremely expensive (we know this!)
- it is fairly easy, in a modern fab, to set a machine to only use PART of a mask set, and when stepping, to move the wafer based on the subarea of the mask that is used, not the whole mask.
Then imagine we do the following. The full Max mask set includes
- a Fusion area at the very bottom of the die
- two GPU+memory areas at the bottom
- an IO (and similar "one time" stuff. Display controller, ISP, Secure Enclave, etc) area at the top.
Now we can use this single mask set to make multiple different Max's.
- Max Ultra1 has Fusion and IO section
- Max Ultra2 has Fusion but no IO section
- Max Normal has no Fusion
- Max Minus has no Fusion AND is missing a stripe of GPU from the bottom
The details are unclear (and maybe the Max Minus does not exist as I describe it, it's always a fused or yield-salvaged Max Normal) but the geometry seems to lend itself to this idea. And it avoids some (not all, but some of) the "waste" that you are, reasonably, worried about – a machine with all these extra Display Controller and IO ports that don't really make sense. Maybe a future design will figure out a way to pack more "unusable" stuff in the IO area (perhaps two or even three of the Display Engines)?
The wildest version of this idea says: why not just cut off the entire top half of the design (more or less, as far as the memory controllers go). What THAT gives you is a GPU-only chip...
Giving Apple some degree of mix-and-matching for building Ultra's and Extreme's.
Eg, for example, perhaps the first version of the Extreme could be a Max Ultra1 and Max Ultra2 (so looking like a current Ultra pair) along with two Max GPU's. Kinda like a dGPU, but without the dGPU downsides.
Of course this is wild speculation, but this generation has shown us that it's silly to get locked into certain ideas ("a cluster is four cores", "the Pro is a chop of the Max") when alternatives arise that have interesting potential in terms of opening up new possibilities.
Last edited:
The M3 Max has double the encode engines compared to the M3 and M3 Pro... My understanding, though, is that coming from a 2018 Intel base model, any of the M3 family should absolutely fly. That said, I don't do video stuff, so I'd recommend starting a separate thread so you can better get advice from people who do.
Be still, my beating heart!
However, I would be shocked if the M3 Max accommodate a 4-way Extreme setup. Left with a feeling that they used all possible die space with no extra of a second Ultra Fusion-imposter. (I know you didn’t imply that, just thinking out loud about the M3 generation.
Just think out loud:
Maybe it is not related to computational enhancement. Power consumption? I/O alignment to optimize motherboard traces complexity (if a revised die layout resulting in reduction of number board layers- could be a big deal)
some surprises
to me it is clear that apple probably regrets making the m1 and m1max so powerful because even before the m3 release there was little differentiation in performance between m1 --> m2 ---> m3 such that users didnt see enough performance differentiation to be upsold.
I think thats what we're seeing here with this Frankenstein release of less cores less bandwidth, deprioritization of m3 pro with no thunderbolt 4 and capability to use multi-display. all of this is on purpose to be able to upsell users to the m3 max.
the problem remains. unless you begin exporting/importing 4 hours of 4k video and have 50 tabs open, and editing 100 megapixel photos ALL at once, the typical 99.95% of users will see little-no performance difference between M1 chips and m3 chips.
the real question which still remains unresolved is how good is m2pro-max to m3 pro-max GPUs. with so little gaming support it probably doesnt matter. If Apple decides to dive into gaming to take on the switch, we'll probably be on the m5.
so the sweet spot is still m1max to m2 pro-max. theres a reason why apple didnt focus on m2 chip vs m3 chip performance. in the real world it is probably negligible and imperceptible for 99.99% of workloads.
to me it is clear that apple probably regrets making the m1 and m1max so powerful because even before the m3 release there was little differentiation in performance between m1 --> m2 ---> m3 such that users didnt see enough performance differentiation to be upsold.
I think thats what we're seeing here with this Frankenstein release of less cores less bandwidth, deprioritization of m3 pro with no thunderbolt 4 and capability to use multi-display. all of this is on purpose to be able to upsell users to the m3 max.
the problem remains. unless you begin exporting/importing 4 hours of 4k video and have 50 tabs open, and editing 100 megapixel photos ALL at once, the typical 99.95% of users will see little-no performance difference between M1 chips and m3 chips.
the real question which still remains unresolved is how good is m2pro-max to m3 pro-max GPUs. with so little gaming support it probably doesnt matter. If Apple decides to dive into gaming to take on the switch, we'll probably be on the m5.
so the sweet spot is still m1max to m2 pro-max. theres a reason why apple didnt focus on m2 chip vs m3 chip performance. in the real world it is probably negligible and imperceptible for 99.99% of workloads.
Nonsense. The M3 Pro has Thunderbolt 4 and USB 4 just like the M3 Max, and it supports up to two external 6K displays [ https://www.apple.com/macbook-pro/specs/ ]. It also will be a good amount faster than the M2 Pro. Not sure where all of the piteously bleak outlooks are coming from.
So annoying. Our IT will already consider me a snob for wanting a mac, I don’t know what they will think when I want to use 270€ for a single extra cpu core…
Even worse, the max is 3900€ with the 512GB drive.
I get the business logic, but for someone that has zero need for a fast GPU this sets the price of a faster CPU quite high.
I’m borderline thinking of just getting the base M3 model, it’s 370€ cheaper then the one with a single extra P core. With this thinking I’m wondering if they will end up with more buyers downscaling to M3 rather then upgrading to max?
Even worse, the max is 3900€ with the 512GB drive.
I get the business logic, but for someone that has zero need for a fast GPU this sets the price of a faster CPU quite high.
I’m borderline thinking of just getting the base M3 model, it’s 370€ cheaper then the one with a single extra P core. With this thinking I’m wondering if they will end up with more buyers downscaling to M3 rather then upgrading to max?
I agree, that's why I said I wouldn't bet on it. It's simply a plausible possibility, like some of the other things you posted today.
No, the idea of the quad isn't to put more ultrafusion on each die. It's to put the dies in a 4-way cross configuration with a single ultrafusion pointing towards the center. Then you presumably get some sort of traffic director in the middle (though of course it could also do much more, like add cache, I/O, etc.).
No. M3 Pro have TB4. Look at the specs page for god's sake and stop spreading misinformation. The one that doesn't have TB4 is the M3 base variant of the MBP. It replaced the old tb-MBP.
Did you even look at the techs specs? 3D rendering gets a big boost. I guess the simulation and data analysis in sciences gets boost too. It would be nice if someone decides to run big CASA files or big AIPS files. It would seriously hammer the CPU cores hard
I'm not disparaging gaming as a benchmark but why should we have it as a measurement of being good. Gaming doesn't use RAM much but 3D apps use a lot of RAM and can affect the performance by a significant margin.
No. The new sweet spot is gonna be M3 Max and it's on the table for M3 Pro but looking at the preliminary tech specs. It's looking that they can get a little bit better with lesser P cores.
First attempt of annotated Apple M3 processor die shots
www.tomshardware.com
Annotated Apple M3 Processor Die Shots Bring Chip Designs to Life
Colorful highlights and annotations help visualize the family.
Single threaded performance is only up 15% per core but we are down 25% on core count. That means multithreaded workloads (such as code compilation) will suffer.
Furthermore - the people saying M3 pro will match m2 pro in multithreaded performance, that isn’t a a win. That’s a loss, a failure. Multithreaded performance should be going up not flat or down .
When Intel and amd release chips that look to have compromises we are allowed to call them out, Apple should be no different.
The m3 generation GPU looks great, the core counts and tradeoffs are questionable.
So, P-core perf is up 15% and E-core perf is up 30%, and while P-core count is down 25%, it's up 50% on E-core count. With E-cores being between 1/4 and 1/3 the perf of P-cores, MC perf of M3 Pro compared to M2 Pro is between:
((1.15 * 6) + (1.3 * 6 * 1/4)) / (8 + (4 * 1/4)) = 98%
((1.15 * 6) + (1.3 * 6 * 1/3)) / (8 + (4 * 1/3)) = 102%
Eh, hopefully I didn’t screw up my math. Quite sleep deprived currently 😒
Anyway, MC looks like it’ll be about the same, but P-core perf for ST should still be fantastic.
Last edited:
I'm not sure what those calculations are supposed to represent.
The overall performance of the M2 Pro, based on your assumptions, is between 9 and 9.33 M2 P-cores. The M3 Pro gets the performance of 6.9 M2 P-cores from its 6 P-cores and between 1.95 and 2.6 M2 P-cores from its 6 E-cores. The overall performance is between 8.85 and 9.5 M2 P-cores, or between a 2% decrease and a 2% increase relative to the M2 Pro.
Your math is odd... (as JouniS said but I was writing this while their comment appeared)
Assuming an e-core = 1/4 of a p-core:
Let P = the performance of an M2 P-core
M2 Pro = 8P + 4 * 1/4P = 9P
M3 Pro = 6 * 1.15P + 1.3 * 6 * 1/4P = 6.9P + 1.95P = 8.85P
The M3 Pro is lower!
They had an opportunity, with the max going to 12 HP cores, to leave the pro at 8 HP cores and have a nice 4 - 8 - 12 segmentation, instead we have 4 - 6 - 12 which is far less elegant and leaves a big gap between the Pro and the Max.
The M3 Max is a slightly more straightforward upgrade, however, M1 Memory Bandwidth Utilization we know that a single P-core can pull over 50 GB/s of bandwidth, so I also dispute the claims that the memory bandwidth limitations aren't going to crop up with the more limited bus width.