M4+ Chip Generation - Speculation Megathread [MERGED]

[Antony Newman]

Apple's hardware is using considerably lower TDPs than other chips of comparable size. For example, Intel's enthusiast-class CPUs have die size comparable to the Mx Pro chips, yet can consume an order of magnitude more power. Heat dissipation does not seem to be a limiting factor for them. I would say the biggest performance limiter is their desire to build systems that perform well on thermally constrained targets, which requires certain architectural and engineering choices.

For laptops - My understanding is, Apple have as lower temperature limit allowed on the underside of their laptops compared to other laptop makers (partly due to corner case issues that can cause a reaction in some people to long term exposure).

But Apple’s keenness to minimise noise - has resulted in extensive research with teams of psychoacoustic experts and engineers : https://www.powerpage.org/apple-patents-point-to-effort-to-reduce-noise-on-macbook-pro-fan-modules/

The 2013 MacPro (trashcan) was an example where (in my opinion) noise limit - acted as a limiter on the maximum average wattage the device could consume.

The standard M2 Studio utilises a chunky aluminium heatsink; the M2 Ultra replaces this with a (more expensive and heavier) copper version. As Apple could have reduced costs by using the same aluminium design, same heat sink - and simply increased the fan speed - I conclude that there is a maximum continuous Noise level limit that Apple is designing the Mac Studio to.

If there is a ceiling on the noise levels, it follows, that based on Apples current level of system-cooling expertise, there is a maximum wattage for a particular volume. The current M2 Mac Stuidio can dissipate 295W continuously. As there are several ways to improve this by changing the surface to PGS (eg pyrolytic graphite) - or a meta material (to change the local impedance of air - which KEF + B&W lead the field) - I speculate Apple will be ‘forced’ to improve this heat dissipation @ lower white noise levels in their next offering. My finger in the air guess is a 21% increase is on the table if they want to get a license from Panasonic (or if they further sculpt the air channel to minimise turbulence).

 

[leman]

RTX 4070 in the *base* M5??? That seems hopelessly optimistic to me (and I'm pretty optimistic when it comes to AS).

I agree. A 10-core Apple GPU would need to implement dual-issue FP32 per cycle and be clocked 2x higher from what it is clocked now to reach the compute capability of the mobile 4070. Essentially, M5 beinging on par with 4070 means that M5 is as fast as M3 Max. That's a bit much.

If Apple goes for dual-issue FP32, they could double their theoretical FLOPS (and probably increase the real-world performance on compute-limited tasks by 30-50%).

I was thinking, what qualifies as an "order of magnitude" exactly?

What I meant is that enthusiast-class Intel CPUs can draw more than 200 watts in multi-core CPU operation. M3 pro is what 20-30 watts? Maybe 50 watts at peak if you push all the IP blocks.

At any rate, the point of my post was not to compare power efficiency but to point out that other chips routinely run at 5x-10x higher power consumption on comparable workloads. Apple is not even close to thermal limits of current technology.

 

[crazy dave]

I agree. A 10-core Apple GPU would need to implement dual-issue FP32 per cycle and be clocked 2x higher from what it is clocked now to reach the compute capability of the mobile 4070. Essentially, M5 beinging on par with 4070 means that M5 is as fast as M3 Max. That's a bit much.

If Apple goes for dual-issue FP32, they could double their theoretical FLOPS (and probably increase the real-world performance on compute-limited tasks by 30-50%).

And even then it still has to operate in fan-less designs prioritizing battery life. Which also complicates pushing TFLOPs too hard in the base model - even for dual-issue designs.

For laptops - My understanding is, Apple have as lower temperature limit allowed on the underside of their laptops compared to other laptop makers (partly due to corner case issues that can cause a reaction in some people to long term exposure).

But Apple’s keenness to minimise noise - has resuled in extensive research with teams of psychoacoustic experts and engineers : https://www.powerpage.org/apple-patents-point-to-effort-to-reduce-noise-on-macbook-pro-fan-modules/

The 2013 MacPro (trashcan) was an example where (in my opinion) noise limit - acted as a limiter on the maximum average wattage the device could consume.

The standard M2 Studio utilises a chunky aluminium heatsink; the M2 Ultra replaces this with a (more expensive and heavier) copper version. As Apple could have reduced costs by using the same aluminium design, same heat sink - and simply increased the fan speed - I conclude that there is a maximum continuous Noise level that Apple is designing the Mac Studio to.

If there is a ceiling on the noise levels, it follows, that based on Apples expertise, there is a maximum wattage for a particular volume. The current M2 Mac Stuidio can dissipate 295W continuously. As there are several ways to improve this by changing the surface to PGS (eg pyrolytic graphite) - or a meta material (to change the local impedance of air - which KEF + B&W lead the field) - I speculate Apple will be ‘forced’ to improve this heat dissipation @ lower white noise levels in their next offering. My finger in the air guess is a 21% increase is on the table if they want to get a license from Panasonic.

What I meant is that enthusiast-class Intel CPUs can draw more than 200 watts in multi-core CPU operation. M3 pro is what 20-30 watts? Maybe 50 watts at peak if you push all the IP blocks.

At any rate, the point of my post was not to compare power efficiency but to point out that other chips routinely run at 5x-10x higher power consumption on comparable workloads. Apple is not even close to thermal limits of current technology.

Aye it's not even clear what Apple's fan noise ceiling might be in their current devices (the M3 Max is louder than the the M1 Max), what tweaks are available to them in their current Studio designs (I suspect they learned from the Mac Pro cylinder where the trash can chassis had limited TDP adaptability which constrained future designs), or even if they ever plan on offering a Mac Pro-specific chip with higher consumption (maybe still unlikely but a lot more likely now with the apparent plan for Apple to be its own best customer here).

It's thus way too early to say that Apple is constrained on the high end by the TDP of the current M2 Ultra chips. Yes Apple likes quieter machines than the competition but as @leman points out they have a long way to go before they are anywhere near the same league of power consumption and heat removal that Intel/AMD/Nvidia expect. I strongly suspect they have built-in some extra headroom for themselves in their desktop chassis even while prioritzing efficiency up and down their lineup.

 

[DrWojtek]

RTX 4070 in the *base* M5??? That seems hopelessly optimistic to me (and I'm pretty optimistic when it comes to AS).


I was thinking, what qualifies as an "order of magnitude" exactly? If you think about it as a statement about log10() - and I don't think anything else makes sense - then anything from 5x to 50x is going to be one order of magnitude, 50x-500x is two orders, etc. So given that, the statement... may be true, depeneding on exactly what you're measuring, but it's not a slam dunk.

Either way though, I agree, leman's larger point stands.

AMD are about (early 2025) to release an APU with graphics equivalent to RTX 4070. https://videocardz.com/newz/alleged...-very-first-benchmark-features-5-36-ghz-clock

My hopes are that Apple, thanks to M5 being released later and possibly being on 2nm (vs 4nm AMD) and assuming they have some magic sauce (which by no means is confirmed) will be able to do this, or at least come close.

Things are moving fast in GPU again, nVidias improvements since 2000 series is mind blowing. When 5000 series gets released, the 4070 will not be upper range anymore. Allegedly AMD have finally found how to improve RT performance by 3x for their next gen, maybe Apple have some tricks up their sleeves too.

 

[MayaUser]

AMD are about (early 2025) to release an APU with graphics equivalent to RTX 4070. https://videocardz.com/newz/alleged...-very-first-benchmark-features-5-36-ghz-clock

My hopes are that Apple, thanks to M5 being released later and possibly being on 2nm (vs 4nm AMD) and assuming they have some magic sauce (which by no means is confirmed) will be able to do this, or at least come close.

Things are moving fast in GPU again, nVidias improvements since 2000 series is mind blowing. When 5000 series gets released, the 4070 will not be upper range anymore. Allegedly AMD have finally found how to improve RT performance by 3x for their next gen, maybe Apple have some tricks up their sleeves too.

As far as i see that Amd M5 equivalent is around GTX 1650, the 4070 AMD equivalent will draw 120W so that cannot compete with M5 into an fanless design, probably with an M5 Max based on W
So bsed M5 i also suspect cannot reach 4070 level of performance

 

[thenewperson]

AMD are about (early 2025) to release an APU with graphics equivalent to RTX 4070. https://videocardz.com/newz/alleged...-very-first-benchmark-features-5-36-ghz-clock

That’s not their base model though? I can’t really see why Apple would move to compete with that (or the 4070) with the base M5. Maybe the Pro will step up to that, but I’m not too hopeful.

 

[MayaUser]

That’s not their base model though? I can’t really see why Apple would move to compete with that (or the 4070) with the base M5. Maybe the Pro will step up to that, but I’m not too hopeful.

We will see what M4 pro can do...but M5 Pro can reach that level...the base M5 no, no matter what..only if Apple choice no more fanless design mac SoC

 

[Boil]

• M5 Ultra = 32-core Threadripper CPU & RTX 5090 GPU
• M5 Extreme = 64-core Threadripper CPU & Dual RTX 5090 GPUs
• Apple ASi Mac Pro Cube
• ;^p

 

[leman]

AMD are about (early 2025) to release an APU with graphics equivalent to RTX 4070. https://videocardz.com/newz/alleged...-very-first-benchmark-features-5-36-ghz-clock

That is not a M-series equivalent chip though, its more similar to the Max-series. Based on the specs it should be comparable to M3 Max. I don’t see Strix getting close to 4070.

Things are moving fast in GPU again, nVidias improvements since 2000 series is mind blowing. When 5000 series gets released, the 4070 will not be upper range anymore. Allegedly AMD have finally found how to improve RT performance by 3x for their next gen, maybe Apple have some tricks up their sleeves too.

AMD and Nvidia velocity in the GPU space has slowed down considerably in the past few years. Nvidia is pretty much at their peak, they will probably need a new design to make meaningful improvements. AMD just seems stuck. Im not surprised to hear they find ways to boost RT performance, their RT implementation was as basic as it gets.

 

[Pressure]

AMD are about (early 2025) to release an APU with graphics equivalent to RTX 4070. https://videocardz.com/newz/alleged...-very-first-benchmark-features-5-36-ghz-clock

My hopes are that Apple, thanks to M5 being released later and possibly being on 2nm (vs 4nm AMD) and assuming they have some magic sauce (which by no means is confirmed) will be able to do this, or at least come close.

Things are moving fast in GPU again, nVidias improvements since 2000 series is mind blowing. When 5000 series gets released, the 4070 will not be upper range anymore. Allegedly AMD have finally found how to improve RT performance by 3x for their next gen, maybe Apple have some tricks up their sleeves too.

That's just baseless speculation.

If AMD could really make an APU that could match the GeForce RTX 4070 at 120 Watt TDP, that would be the biggest generational leap ever for them.

Right now AMD needs the Radeon RX 7800 XT with 60 CUs to match that at 263 Watt TDP. Going from RDNA 3 to RDNA 3.5 is not that kind of upgrade. Also the memory bandwidth just isn't there either, even with LPDDR5X at 8000MT/s.

 

[DrWojtek]

We will see what M4 pro can do...but M5 Pro can reach that level...the base M5 no, no matter what..only if Apple choice no more fanless design mac SoC

The mac mini

 

[Confused-User]

The mac mini

The mini is not the target for the M5 or any base M chip. The primary target is the MBAir. Secondary targets are the low-end MBP and the iPad. The mini is a distant third, obviously, due to sales numbers... but also because it's not important to pump power into the Mx for the sake of the Mini, when you can also get the mini with the MxPro.

Especially now that the QC SXE is shipping, it's apparent that Apple's uncontested rule over the fanless design is a huge strength. They won't compromise that.

 

[crazy dave]

That’s not their base model though? I can’t really see why Apple would move to compete with that (or the 4070) with the base M5. Maybe the Pro will step up to that, but I’m not too hopeful.

Agreed. The M5 Pro GPU getting to the 4070 mobile is ... possible and even that would be hard. Doable but hard. 4070 Max-Q is much more plausible. It depends if Apple does what @leman suspects they will and doubles up on FP32 units per core (turning one of the Int32 or FP16 units into a FP32/x unit like Nvidia does) and what performance gains they get. 4070 desktop, especially for anything remotely computational as opposed to raster, is out of the question almost no matter what at the Mx Pro's size and TDP. The Strix Halo AMD chip isn't likely hitting that either though at 40 cores should be giving the M3 Max a run for its money in certain scenarios. An M5 Max though? Maybe. Again it depends on what Apple is able to achieve (also whether or not Apple's upcoming Brava and Hidra dies are M5 or M4!).

However, I will say that the existence of Strix Halo and potentially other such large APUs coming, maybe from Nvidia if rumors are to be believed, is why I believe Apple need to keep moving as fast as possible in this space. I'd especially like to see higher MatMul added to the GPUs ASAP. Even if the AI market crashes, having a comparatively cheap (weird to describe an Apple product like that!) AI training device (it's already good at non-batched inference and MatMul in the GPU has other uses) before anyone else could be a significant leg up. In fairness to AMD, Apple achieved the vision AMD had been touting for years, but never had the resources to pull off.

 

[iPadified]

Apples focus on low noise computers is not so surprising. Noice is very stressful and at least for me it affects my efficacy. I ran some local bioinformatics analysis on my iMac 2020 as I could not update my MBP M1 Pro RAM. The iMac2020 sounds like a hair dryer and the analysis takes hours so it is a pain. The probably equally fast MBP M1 Pro is nearly silent under max loads such as 3D rendering and that silence is much more pleasant and will not affect my efficacy doing other stuff. My 2013 Mac Pro (unfortunately broken) would have been pleasantly silent as well.

Personally I think Apple is on the correct track and puts a noise ceiling for desktop computers to not affect the work environment. Not max performance in terms of compute but a OK performance in pleasant environment. Mac Pro is under the desk computer and that is another situation altogether.

 

[streetfunk]

Not max performance in terms of compute but a OK performance in pleasant environment.

i like that sentence

 

[tenthousandthings]

Some basic thoughts about the roadmap. We are, I think, far enough into the Apple silicon on macOS era to start drawing conclusions. We'll know more after M4 Pro and beyond rolls out, but for now:

[1] The closest thing we have to a roadmap is TSMC's, which evolves behind the scenes and is set in stone when it reaches the public.

[2] We know it takes 1.5 to 2 years for TSMC's customers to go from initial tape-out to production on their leading-edge nodes. TSMC's leadership has stated this more than once, most recently in the earnings call on July 18 (see p. 18 of the transcript cited below).

So let's assume M4 Pro on N3E is going to launch in October 2024, with volume production having begun in Q4 2023. So the M4 Pro was finalized in H2 2021 or H1 2022.

Think about that. The engineering decisions Apple is making for N2 products launching in late 2026 have already been made. The architects are currently working on 2027, and making decisions about tape-outs on N2P and/or A16. TSMC hasn't released all of that roadmap (especially with regard to the timeline for N2P), but enough of it is official that we know A16 with Super Power Rail is set to go into production in 2H 2026, with Apple product launch in 2027. M7, here we come?

M1 Pro/Max (2021) = N5 (2020)
M2 (2022) = N5P (2021)
M3 (2023) = N3 (H2 2022)
M4 (2024) = N3E (H2 2023)
M5 (2025) = N3P (H2 2024)
M6 (2026) = N2 (H2 2025)
M7 (2027) = A16 (H2 2026)

[3] You can see in that earnings call that TSMC can't talk about their progress in advanced packaging without leaking their customers' plans, so they just ignore and/or deflect specific questions about it. One person asks about silicon interposers (pp. 12-13, the CEO's answer is "I'm not going to share with you because this is related to my customers' demand."), and another states outright that TSMC has one and only one InFO customer (obviously Apple, for the M1/M2 Ultra, see p. 14, question from Laura Chen). The CEO's response, however, does seem to crack the facade: "... Let me share with you, as my customer moving into 2-nanometer or A16, they all need to probably take in the approach of chiplets. So once you use your chiplets, you have to use in advanced packaging technologies. ... HPC is moving faster because of bandwidth concern, latency, footprint, or all those kind of things. For smartphone customers, they need to pay more attention to the footprint as well as the functionality increase. So you observe my big customers taking the InFO first and then for a few years and nobody catch it up, they are catching up. ..."

[4] So it seems like there's a decent chance there could be a fork in the road in 2027, where N2P is used for monolithic smartphone SoCs and A16 is used for heterolithic (probably not the right word, or not even a word!) HPC chips that use advanced packaging...

The above quotes are from this: Edited Transcript July 18, 2024

 

[PaulD-UK]

"So the M4 Pro was finalized in H2 2021 or H1 2022."
In H2 2021 or H1 2022 artificial intelligence to Apple meant Siri, Apple Car self driving or further perfecting the look of backlit woollen sweaters with the iPhone camera....

Surely that means any current talk about the M4 being an AI Great Leap Forward is mostly hype, with a stock split being done to double the neural engine benchmarks?

 

[jdb8167]

"So the M4 Pro was finalized in H2 2021 or H1 2022."
In H2 2021 or H1 2022 artificial intelligence to Apple meant Siri, Apple Car self driving or further perfecting the look of backlit woollen sweaters with the iPhone camera....

Surely that means any current talk about the M4 being an AI Great Leap Forward is mostly hype, with a stock split being done to double the neural engine benchmarks?

The A14 Bionic was released in 2020 with a neural engine. Same with the M1. Apple has been planning and using machine learning for a long time already (since the A11 Bionic in 2017). Given the above timeframes that means that Apple was designing their A14/M1 NPU between 2018 to 2019.

 

[PaulD-UK]

"Apple has been planning and using machine learning for a long time already...
Apple was designing their NPU between 2018 to 2019."

For the aforesaid AppleCar and iPhone camera performance, surely?
Quite a different target?

 

[jdb8167]

"Apple has been planning and using machine learning for a long time already...
Apple was designing their NPU between 2018 to 2019."

For the aforesaid AppleCar and iPhone camera performance, surely?
Quite a different target?

Not sure why you think that? It's been used in iPhones and iPads for many different tasks and now in the Apple silicon M SoCs as well. If Apple was planning a self-driving vehicle then they would need a much more powerful NPU than what was in the 2020 SoCs and most of the camera processing is probably done on the separate image processor.

 

[PaulD-UK]

"If Apple was planning a self-driving vehicle then they would need a much more powerful NPU..."

Wasn't the planned Car release date 2025+?
That's my point.
Surely the M4's NPU was designed for reasons far removed from what today's AI hype is all about?
The AI Great Leap Forward will be more like 2026+?

 

[Populus]

"If Apple was planning a self-driving vehicle then they would need a much more powerful NPU..."

Wasn't the planned Car release date 2025+?
That's my point.
Surely the M4's NPU was designed for reasons far removed from what today's AI hype is all about?
The AI Great Leap Forward will be more like 2026+?

Yes, I expect the big leap forward for on device AI will likely come with the A20 or M6 gen, circa 2026, with a bump not only in Neural Engine cores, but also with bigger and faster on-chip RAM modules.

 

[jdb8167]

Surely the M4's NPU was designed for reasons far removed from what today's AI hype is all about?
The AI Great Leap Forward will be more like 2026+?

Yes, I expect the big leap forward for on device AI will likely come with the A20 or M6 gen, circa 2026, with a bump not only in Neural Engine cores, but also with bigger and faster on-chip RAM modules.

Curious what great AI leap you are envisioning? The machine learning stuff that Apple is putting in the next macOS doesn't seem to need anything more than an M1. Apple is unlikely to be training large language models on device anytime soon.

 

[Confused-User]

[1] The closest thing we have to a roadmap is TSMC's, which evolves behind the scenes and is set in stone when it reaches the public.
[...]

So let's assume M4 Pro on N3E is going to launch in October 2024, with volume production having begun in Q4 2023. So the M4 Pro was finalized in H2 2021 or H1 2022.
[...]

[4] So it seems like there's a decent chance there could be a fork in the road in 2027, where N2P is used for monolithic smartphone SoCs and A16 is used for heterolithic (probably not the right word, or not even a word!) HPC chips that use advanced packaging...

I think you make some very good points, but I want to take issue with a few specific things quoted here.

TSMC's roadmap isn't set in stone. It never is because they aren't the final arbiter, physics and engineering are, and it's not done until it's done. They have a fairly good record, but they have famously screwed up once recently (N3B rollout delay), and revised a released roadmap recently (moving BSPD down the road a bit). None of this really impacts your main point, though.

I don't see why you think volume production for the M4Pro started late *last year*. That seems wildly unlikely. I think if they were allocating wafers to the M4Pro they'd release it (along with the Max) earlier. More likely they went with the base M4 only to reduce the impact of defects as they ramp. Unless you just mean that N3E started volume production last year, not the M4Pro?

Your "fork in the road" prediction is interesting but I think you may be reading too much into the TSMC comments you quoted. Advanced packaging may reduce area substantially while raising z-height a bit. I think they can afford some z if it buys them area, but there's always heat dissipation to consider. So it's not at all clear to me how that's going to play out.

 

[leman]

with a stock split being done to double the neural engine benchmarks?

I am curious what you mean by this?