I know some of the differences will be better efficiency and less heat for the 3nm chips but what else is there? The reason I am asking is that I am wondering if there is any logical advantage to waiting three months for a 16" MacBook Pro Max (January) after Apple makes the switch to 3nm chips.
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.
What are the differences between the 3nm and 5nm M2 Max CPUs?
- Thread starter Amazon Rainforest
- Start date
- Sort by reaction score
Well no one actually know what is coming out yet. I highly doubt Apple would split architecture on one line of chips. I would think that Max chips will still be on 5nm. M2 line no matter, regular, pro, max, etc. will all be on 5nm. M3 would be on 3nm and all successive chips. Maybe I am wrong and I have heard the rumors but it doesn't make sense nor follow Apple's market strategy of maximizing profits. They will get as much as they can out of 5nm and next year will be m3 with 3nm. Otherwise I don't see 2nm or 1nm coming any time soon as 3nm is still not even released so it is going to take time for the new process to be developed.
What I don't understand is why Apple didn't use 4nm on M2 and 3nm on M3? TSMC and Samsung already produce 4nm and I am sure it would be better than the newer 5nm process at TSMC unless Apple wants a more significant difference between M1/2 and M3.
What I don't understand is why Apple didn't use 4nm on M2 and 3nm on M3? TSMC and Samsung already produce 4nm and I am sure it would be better than the newer 5nm process at TSMC unless Apple wants a more significant difference between M1/2 and M3.
The main advantages of a node shrink are what you already metioned: higher performance and/or less power They can use the process improvement all for efficiency, all for performance, or a combination of both. And if they're limited in the size of the die they can use, a process shrink allows them to put more transitors on the die, which allows them to expand upon or add to processing components (number of cpu/gpu cores, size of coprocessors like the neural engine, video deocoder, etc.)
But: (a) we have no idea how Apple would make use of a node shrink; and (b) we have no idea if there will be a node shrink on the next gen of MBP's; they could still be 5 nm.
Having said that, I'd wait regardless, simply because we're near the end of the product cycle for the 14"/16" MBP's. Even if they don't go to 3 nm, you'll still see improvements from the new M2 microarchitecture. There may be other new features as well. And if those don't matter to you, you'll be able to get the current models at a discount.
But: (a) we have no idea how Apple would make use of a node shrink; and (b) we have no idea if there will be a node shrink on the next gen of MBP's; they could still be 5 nm.
Having said that, I'd wait regardless, simply because we're near the end of the product cycle for the 14"/16" MBP's. Even if they don't go to 3 nm, you'll still see improvements from the new M2 microarchitecture. There may be other new features as well. And if those don't matter to you, you'll be able to get the current models at a discount.
Last edited:
Nothing they could add using the N5 process. N3 gives them a bit more transistors to play with at the same die area, that’s all.
Don’t they already have two kinds of “tensor cores”?
How do the reticle sizes currently compare between N5P and N3? I did some research awhile ago, but couldn’t find much about that in particular. IIRC, M1 Max was pretty close to the N5 reticle limit.
Edit: M1 Max was 432 mm^2 according to Wikipedia, and in August 2020 techradar claimed the DUV and EUV reticle limit was a whopping 858 mm^2. So, maybe not close to the limit after all?
Last edited:
Don't know, honestly. I just included that as a possible effect, for completeness. There's a discussion here hypothesizing that, if they wanted to add hardware ray tracing to the M2 Max, that might push them up against the reticle limit with N5: https://talkedabout.com/threads/apple-m1-vs-m2.3135/
...That got me wondering about how they make the Ultra. From this discussion, it looks like only each of its component Max chips need to fall below the reticle size; but they don't sound like they know definitively: https://semiwiki.com/forum/index.php?threads/apple’s-new-m1-ultra-packs-a-revolutionary-gpu.15637/page-2
Last edited:
There is not a chance Apple is going to start a new production process for the A15. It will all be on 5nm.
Most likely, the iPhone will get 3nm first with the A17 and then the Mac’s and iPad’s will follow with the 3nm M3 being based on the 3nm A17.
Tim Cook is all about operational efficiency, since he is an Operations guy. So all these rumours about a 3nm M2 Pro and M2 Max seems very unlikely to me.
I share your doubt, but I also think it might be a justifiable risk to postpone the release of the new Pros a bit (until spring 2023) if they can get 3nm tech, as this most likely will mean undisputed performance superiority.
I agree. Even better would be to relase the MBP's with 5 nm M2's in the fall, and then give it a spec bump in the spring with 3 nm M2's. That would allow them to keep the MBP's on the cutting edge w/o delaying their next update until the spring. I don't know if they'll do this (they may decide it's too resource-intensive), but we shall see....
I don't think this will be a good plan. First, it would likely be prohibitively expensive to set up mass production of two series of chips for such a short duration. Second, users don't like this kind of stuff — at all. There will be a lot of complains and bad publicity. Just remember what happened in the past, e.g. when Apple introduced the 6600M...
Knowing Apple, I think it's fairly safe to assume that they have two chip series in development and "ready to go" — one for 5nm and one for 3nm — and they will choose to mass produce only one of it as the M2 Pro/Max. They will decide for the 5nm variant when it is judged that the risk of 5nm production is too high, and they will decide for 3nm if they are confident that they can roll it out in sufficient quantities soon enough. From both the financial and PR standpoint it is better to delay the Pros until the new chips are ready — the current M1 crop is still very good and the users won't mind if the new machines arrive a bit later. In fact, it might prompt some of the current M1 users to upgrade.
I am quite confident that we will eventually get to one new release for the entire Mac line each year. But it will take some time to get there. At the scale Apple operates logistics is the crippling factor.
I was thinking the same thing, and said pretty much that a few days ago in another thread:
[My thinking was that Apple wouldn't be updating the MBP's any more frequently than yearly and, with supply chain issues, the gap will likely be even larger for the foreeable future—thus they'd hamstring themselves by releasing 5 nm MBP's in the fall.]
However, I changed my mind because a poster replying to me pointed out that Apple indeed used to do much more frequent updates (twice yearly, in fact) of the MBP's, and people liked that the MBP was was being continually updated to keep up with technology. I think what people don't like is not knowing. If Apple returned to a more frequent update cycle for the Mac, and people expected that, it would be a plus, not a minus. That way no one gets caught by surprise when Apple updates the computer you just bought six months later.
Thus the only downside to more frequent updates is the resources it demands from Apple, and I'd already mentioned that ("they may decide it's too resource-intensive").
Here's a very nice table from Wikipedia showing the (on average) semiannual updates we used to get.
Agreed, though while I also expect they'd like to get on an annual cycle (particularly for their microarch. updates, to synchronize with the iPhone, I think it will be staggered, as I've mentioned earlier, and for the reasons I explain below:
Last edited:
That is very true, however, these semi-annual updates were just a minor spec bumps using the same technology platform. You'd maybe get a 100-200mh CPU bumps, that resulted in maybe 3-5% performance improvement (if at all), mostly an encouragement to customers who are on the fence of buying a Mac mid upgrade cycle.
None of these spec bumps cost anything to Apple, they just got the new party of chips from Intel and put them into the logic board. I kind of doubt that it even costed anything to Intel, since these were most likely the same chips just configured with slightly higher frequency. Chip manufacturing getting more cost effective/mature over time probably also paid a role.
But the context we are discussing here is very different. We are talking about a major upgrade and an entirely new chip production line. It's a major investment for Apple to roll out a new family, and producing it for just a couple of month is likely not financially viable. I would also expect a major performance difference between a 5nm and a 3nm chip. IMO, ff Apple even goes back to the semi-annual updates (which is honestly doubt as that was a side-effect from Intel's release schedule), these will be just minor clock boosts again with no real changes in manufacturing process.
Last edited:
The jump from 5 nm to 3 nm would likely require an entirely different design to work. Node shrinks aren't as simple as just applying the same design at a smaller size. For that reason I see it as highly unlikely that Apple simply makes two batches of the same chip at different node sizes. The closest thing we've had to this in the past was the A5 used in the third generation Apple TV that went through a revision with a smaller node size, but that also came with a redesign where one of the CPU cores was removed (and not just deactivated like on the original chip).
That just goes to prove that you can implement the "same" core (in terms of high-level design and name) on different processes. Of course it's not a simple matter of setting the scale in the print dialogue to 60% (/s) - but that doesn't mean it requires an all-new design. Since we're at a time of transition between 5nm and 3nm - with extra uncertainties over timing caused by recent events - it is perfectly plausible that Apple specifically planned the Avalanche/Blizzard-based CPUs as 3nm with a 5nm Plan B in mind. Which would explain the current conflicting rumours...
Apple's chip names - A5, M1, M2 etc. appear to be based on the high level logical design of the CPU cores - ARM Cortex A9 for the A5, Avalanche/Blizzard for M2, Firestorm/Icestorm for the M1. Adding/removing cores or other functionality doesn't seem to affect the stem of the processor name (e.g. the M1 pro/max/ultra have quite different permutations of CPU and GPU cores, a next-gen RAM interface and extra hardware codecs c.f. the regular M1, but they are still "M1" series).
Ultimately, the names of any new Apple Silicon chips will be chosen by Apple's marketing people, not their engineers. They don't even have to stick to the current numbering scheme which, AFAIK, Apple have never publicly defined or committed to. If Apple do release 'stop gap' 5nm M2 Pro/Max chips this year then it could make marketing sense to call the following generation "M3" even if they use the same cores as the M2.
Realistically, though, we currently have only a few data points from which to guess Apple's post-Apple Silicon update cycle - over 18 months for both the MacBook Air and 13" MacBook Pro, and nothing yet for the M1 iMac or Mini. The M1 Pro/Max 14/16" MBPs were only launched last October so if 18-month cycles are going to be typical there's no reason - apart from sketchy rumours - to expect M2 Pro/Max versions until next spring. That's getting into the time frame where 3nm production might be rolling. There's no particular reason for Apple to release a 5nm stop gap unless (a) they don't expect 3nm that soon or (b) MBP sales are starting to tank because of the lack of updates. Only Apple know that.
One factor is - M1 started with a hefty head start over Intel (unless you used Intel logic and compared it to a barely portable brick with 2 hour battery life) and there's not currently any real point of comparison between Mac and Intel (i.e. Apple aren't trying to sell machines with gen 12 Intel chips when Dell are advertising gen 14) so there's no real rush - besides MacRumors pundits - to release new machines just yet. Especially the higher-end MBPs which you wouldn't expect to be as fashion-accessory-y as the MBA and which aren't currently due for a physical redesign.
Apple will likely discontinue all M1 models when the Pro's move to M2. The only reason the M1 air still exists is due to inventory of M1 components.
That and that they can continue to have a model under $1000. Don’t ignore the power of price points.
The fixed costs of manufacturing a chip on a leading edge node are enormous. The design itself is a minor part of it. Producing the masks is fantastically expensive. It only makes sense to bring an existing design to a new node if you are selling a huge number of them. I don't think any Mac SoC family is doing those kind of volumes. The iPhone probably does enough volume, but the product line advances too quickly to make a node change make sense.
How much change would be required of designs moving from N3 to N3E or N3P or N3X (same sized nodes)? Might those types of changes make it easier for Apple and TSMC to jointly build in some schedule flexibility around things like yields and costs?
The M1 Max was used to constructed the M1 Ultra. 2 *432 --> 864 is hovering around the reticle limit. The packaging technology Apple used for the M1 Ultra ( Info-LSI) does have a 1x reticle limit. Apple could possible go to a path where they abandon Info-LSI ( the quad die "Extreme/much-bigger-package" is likely to blow far past what Info-LSI can do even with some improvements. ) or a 1.25-1.5x reticle Info-LSI shows up.
Or it could be that much higher clocks (performance ) or much lower power is what the main levers on N3 would be. Going to a "Max-like" die that is 406mm^2 would:
1. allow Apple to stay under the reticle limit for the current Info-LSI packaging. Basically simpler, more cost effective packaging.
2. allow Apple to get more useful dies out of single wafer. Smaller dies have incrementally higher yields. They also by definition just take up less space on the wafer. That would mean they would need less wafers to make more Ultras and Quads. Given Apple is *still* behind on getting Studios out the door having more supply might actually help. N3 wafers also cost more. So more useful dies out of a wafer lowers the per unit cost assigned to each working wafer. Again is works out right saves Apple more money which they can use for bigger margins.
What Apple pays TSMCs for is putting wafers through the fab process. How many working dies come out the other side is not TSMC's problem. You paid for the wafer you get the wafer. If there are 4 or 40 or 400 working dies on the other side that is your (Apple's ) cost problem. Yes, TSMC needs to provide decent yields to get customers to order a reasonable number of wafers to be process. So they care , but the actual costs is falling on the customers. And the customers control how big the dies are. TSMC largely doesn't dictate that to the customers. If customer makes a bigger die, then the yield goes down.
If Apple stays on N5P for a "More cores and more stuff" M2 Max then for a 8% bloat ( 466mm^2 ) they'll get less dies and lower yield out of a N5P wafer than they were getting before. N3 wafers may cost sufficiently more that trade-off is worth it. Or perhaps not.
3. allow Apple to mitigate risks associated with the new tech of N3. N3 has a "Flex Fin" technology that really hasn't been tried in large scale production before by any fab. Implementing a "working" M2 architecture on top of a new fab process would be pretty close to following Intel's old "tick/tock" strategy. Either do large tweak to the microarchitecture or do a large tweak to the fab process , but largely avoid doing both at the same time.
There is lots of hype about the first N3 implementation will lead to some massive leap in core counts or power savings. Perhaps that is much more narrow. No massive core count increases (take same increase as bloating the N5P would give) and the power savings also incremental (as fell out the new technology without overly optimizing to save time. ).
406mm^2 would be in reach if just waited for N4P (-6%) . A conservative use of N3 (-20%) could take the size down to 346mm^2. That's still 100mm^2 bigger than the current M1 Pro. It would still be a 'mid-size" die, but far closer to the size of chips Apple would likely prefer to make ( more volume out of fewer wafers and better margins. )
And 4 * 346 is 1394mm^2. That is relatively very large, but more tractable than 1864mm^2. Again bigger , more complex packages likely lead to incrementally lower yields (and higher costs... which Apple likely doesn't want to pay.).
As for Apple would never do that.
A10 -- 16nm 125mm^2
A10X -- 10nm 96.4mm^2
Yes , the iPad Pro chip was smaller than the iPhone chip. And 10nm was also a fab tech inflection point for TSMC.
The M2 Max bloating on N5P would not matter as much if Apple was decoupling the Ultra from using exactly the laptop Max die in the desktops. The Ultra and bigger are never used in a laptop. If Apple stripped the UltraFusion logic and connectors from the die they'd get a smaller die or could swap that area for a few extra E-core or G-core space. Even more so for a M2 Pro since it didn't have the often completely unused ultrafusion connector soaking up space anyway.
So if Apple couples the MBP 14"/16" Max dies to solving the unit volume problems of the Studio and more so the Mac Pro then there is a decent chance the Max could get pulled into N3 because it is a better fit for those desktop systems. ( probably get some decent reduction in power consumption to help the laptop also as somewhat a side effect. )
If Apple decouples the laptop Max die from the desktop issues then N5P could make some sense for M2 Max. M2 Pro doesn't present thermal growth problems since same current systems have a Max to deal with. (and a M2 Pro Mini has gobs of headroom because built for Intel desktop set up. )
Last edited:
AMD makes Ryzen 7000 official: Launching September 27, starting at $299
Company promises a big 29% jump in single-core CPU performance from Ryzen 5000.
arstechnica.com
Funny how Zen4 can span N5 and N4 while Apple can't possibly do M2 on any other node than N5P. Or that Zen5 can span N4 and N3 while Apple "has to be stuck" on N5P for the rest of the M2 line up because ... can't be done.
Or that the A10 was 16nm and the A10X was 10nm because Apple can't split an implementation over multiple fab processes.
If Apple (or AMD or Intel) plans out at the start 3-4 years in advance to put a architecture on two different fab nodes it is entirely possible. Folks have done it. The old , highly successful Intel 'tick / tock ' strategy was largely built on that premise.
Doing it ad hoc , 'on the fly' is costly and unproductive. Doing it as part of a long term plan can work quite well.
Probably, typo but there is no M1 coming next year. The M1 line up is done; Apple stated as much.
Getting a new MBP 16" out this year or next is largely what the debate boils down to . Apple needed relatively high volume unit delivery inside of 2022 then N3 isn't it. If Apple's plan was to wait over a year to update the MBP 16" then N3 is viable.
There is lots of commentary about Apple "has to" release Macs on 12 month intervals. They don't need to do that. They haven't done that in the M-series product line up at all. Largely didn't do it with iPad Pro either.
The notion of "keep up with the every September iPhones because Apple 'cares' " isn't a solid foundation. That fixed in stone deadline Apple has hung around the iPhone seems about to finally crack ( which was somewhat inevitable.). The Pro models getting A16 and the 'regular' models getting A15 is demonstrative that it isn't a good model to slavishly follow.
Because Apple doesn't do the kind of volumes AMD does for its regular (chiplet) cores. Not to mention AMD can mix & match their chiplets in almost any combination now, the image you're seeing is for zen4, zen4c & zen4 (x3d) cores. It doesn't say which cores will be made where, besides 4nm is just an optical shrink like they did with 14-12nm. So no Apple will not be making (monolithic)chips on two separate nodes, in fact no one does this except AMD & even they'll do it very rarely like zen3 based 6xxx mobile chips.
It's H1 - first half next year. I don't follow Apple release schedules as closely but that's when things should get more clear, if not this year then probably next year.