What relevance is that? The M3 has already been designed at this point, so superseding it quickly would be expensive on the design side? What point are you trying to make, please?
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.
Apple Debuts Next-Generation M4 Chip
- Thread starter MacRumors
- Start date
- Sort by reaction score
So, M4 vs M3, the M4 has:
A better display engine.
Better GPU, the M3 marketing slides mentions everything from 20% to 2.5x up from M2, so I'm assuming the 2.5x is related to ray trancing. Which begs the question if the 4x is "just" better ray tracing HW or something else?
More E cores, and probably some other CPU related improvements. Since the M3 was 20% better than the M2, again according to marketing materials, and while 10 cores is 25 % more than 8, and 1.2*1.25 = 1.5. I really, really doubt that's what they mean.
Better NPU, they probably started supporting lower precision ops, which is where a lot of that boost comes from. Which is fine, as inference doesn't need anymore anyhow. If a bit sneaky.
Better memory BW. This may or may not translate to more usable BW (as per the M1 Pro/Max), but I'd be surprised if it wasn't (usable).
Less power usage.
A better display engine.
Better GPU, the M3 marketing slides mentions everything from 20% to 2.5x up from M2, so I'm assuming the 2.5x is related to ray trancing. Which begs the question if the 4x is "just" better ray tracing HW or something else?
More E cores, and probably some other CPU related improvements. Since the M3 was 20% better than the M2, again according to marketing materials, and while 10 cores is 25 % more than 8, and 1.2*1.25 = 1.5. I really, really doubt that's what they mean.
Better NPU, they probably started supporting lower precision ops, which is where a lot of that boost comes from. Which is fine, as inference doesn't need anymore anyhow. If a bit sneaky.
Better memory BW. This may or may not translate to more usable BW (as per the M1 Pro/Max), but I'd be surprised if it wasn't (usable).
Less power usage.
Last edited:
The point is don't only look at direct manufacturing cost and keep in mind that the decision to switch nodes, had to be done at least 2 years ago.
But the other poster stated they made the switch to a cheaper node from a more expensive node? Are you agreeing or disagreeing with them? We know it's expensive to design chips, but the discussion is about saving money on build costs.
it'll go to waste on the iPad thanks to iPadOS. hopefully 18 is more than just AI or maybe we need to wait for 19 and keep hoping the make an os that can actually take advantage of the hardware.
it'll be interesting to see M4 in mbp, iMac, studio where it will be able to get pushed hard.
it'll be interesting to see M4 in mbp, iMac, studio where it will be able to get pushed hard.
HobeSoundDarryl
macrumors G5
Ultra is apparently no longer "linked" to MAX (no longer twin MAX chips linked together)... so nothing stops it from rolling out whenever it is ready... even before PRO & MAX.
Apple being Apple and profit mattering more than anything else, I would expect them to roll out M4 Ultra Macs ASAP and basically evolve the releases so that ULTRA leads instead of rolling in LAST... only to be somewhat-to-mostly superseded by the next gen MAX only a few months later, Roll out ULTRA first and those who basically die for "most power" and/or "highest number painted on a chip inside" (that they'll never actually see) may pay up for ULTRA instead of waiting on "weaker" variants to come in a few months. Furthermore, ULTRA would then be "king of the hill" for up to the duration of the M-X generation instead of holding that crown for only a few months.
Why would nearly anyone buy an M2 ULTRA-based Mac since M3-MAX hit? Why would anyone buy one now? But roll out M4 Ultra BEFORE MAX and that question gets answered very differently... and for the duration of the chip generation. My guess is Apple choose profit-per-unit maximization and flips the old apparent pattern... and I'd guess that will start very soon.
Same here. I will play with both devices hand-in-hand and make a final decision one week later.
Yes. Now they're going to take 0,0001s longer for takes 99,9% overall tasks
No thermal issue at all.
Competitors are catching up on NPU performance. Apple needed to take the lead on this.
Details matter, Jim. It's "iPad". Small "i", big "P".
And MacBook Airs are not suddenly "slow". That's exaggeration. It's just that the iPad are "faster" than "fast".
Yes. M4 is a M3 with faster NPU performance.
Samsung is all over the place saying that S24 has an IA hardware. Now, Apple improved its NPU engine to say "mine are faster than yours".
It takes years to roll out new silicon. I'd be surprised if Apple has an internal cadence lower than 2 years.
Was this known a year ago? Then they might have had the time to do something. If not, then no. (Also note that the exploit being published isn't the same thing as manufacturers getting to know about it, most likely they get that information beforehand).
They roughly doubled the number of operations per second compared to the M2/M3. That’s not bad at all, but also not a major game changer with regard to AI computational power.
Any chance this means we'll get a Studio Display at 120hz, or better? Any chance for future Studio Display using display port? Thanks!
Not necessarily. They were careful to specify that the M4 provides the same computational power as the M2 at half the power usage. Given the reduced battery capacity (reduced thickness), they are probably favoring power efficiency over raw speed for the M4s on the iPad.
Releasing M4 ahead of the perceived schedule makes sense. Lower demand for iPad gives them time to get yields up for release in Mac. If it’s a chip related to iPhone there’s benefit there too
False. Even if we had no hard data, it would be utterly unimaginable that they'd go the to the trouble of entirely redoing layout (as is necessary to go from N3B to N3E) without making any improvements. If nothing else, the choices you get from using FinFlex open up huge possibilities for optimization (of either power or performance).
But we do have hard data.
If you look at Apple's page on the M4 you can see that they did additional design work on the CPU and GPU. They call out several things - the CPU is even wider (!!), the GPU has mesh shading hardware, etc. There appears to be something new in the CPU for ML - NOT the NPU changes - so perhaps that's an AMX upgrade?
The memory controllers are upgraded, almost certainly to support LPDDR5x. The display controllers are upgraded (I'm going to go out on a limb and guess these got REALLY extensive changes, thanks to FinFlex, beyond the new support for the "tandem" OLED).
There's *so much* we don't know yet, too. Is the NPU meaningfully updated from the M3? Or are the numbers mostly due to changing from measuring FP16 to INT8? It's definitely going to be better, but we don't know how much. What other GPU and CPU changes are there? Has the uncore gotten further improvements, especially to the NoC? What's the clock speed?
No. M3 is superior to M2 iso-performance in every way.
They couldn't make them the same even if they wanted to, due to the incompatible process, and also not even close - see above.
Probably, though I expect Apple knew about this long before it went public.
But GoFetch is not a practical exploit. It's not something anyone needs to worry about in real life, because the threat model is impractical. If an adversary can exploit it, they will have far easier ways to steal information, almost definitely including keys.
Yes, it should be mitigated in hardware, to prevent further cascading issues over time and with better processors. But for current hardware, it's meaningless unless you're a security researcher.
It's likely that they're saving some money, but not likely that it's significant. The M4 has 12% more transistors. Apple got a favorable price on N3 because nobody else bought it, so the chances that the cost difference (given the size difference) is more than 10% or so seems low to me. However... only a few people in TSMC and Apple will ever know for sure.
Not a good compare- this is in the iPad, with likely lower clocks. Wait for the Mac release.
Soooo wrong. See above.
Doubling is definitely a game changer... at certain breakpoints. 200ms feels much slower than 100ms to humans, and 2s feels much slower than 1s. On the other hand, 10 minutes may not feel that much slower than 5 - both require you to go do other stuff while you're waiting. I think on a practical level, you're wrong- doubling will be significant for a lot of frequent tasks.
The big question is, is this real or is it an artifact of switching between measuring FP16 and INT8? I mean, it's real either way, but if it's not an artifact then it represents a much bigger silicon change.
Close, but that's not quite it. The issue isn't Mac volume. It's iPhone volume. M4 and A18 will use the same silicon, N3E. Apple will already be ramping A18, to be able to deliver iPhone is massive volumes at release date, when they start shipping M4 Macs. Every Mac they ship with M4 reduces (or at least defers) the number of iPhones they can sell, unless/until TSMC gets ahead of demand... which they probably won't until well after the iPhone ships. It looks like they've calculated that they can in fact afford that this year. Finally.
That's possible. We don't know.
Except for the M3 process likely being pricier than the M4's.
It’s not a game changer with respect to LLMs. If they are unusably slow on an M2/M3, they will still be on an M4. They could have increased the number of neural engines for more parallel processing, but they didn’t (would probably draw too much power).
I'm sorry, that's just nonsensical. Somewhere there is a breakpoint between usable and unusable. And there are tasks where the speed of performing that task on the M2 is on one side of that line, while the M4's speed is on the other side of that line. That should be obvious to anyone over ten years old. (There will similarly be a dividing line between "annoying" and "just fine", to which this same reasoning applies.)
The question is, do commonly performed tasks (or ones that we'd like to commonly perform) fall into that category? I don't think there's any way to know that yet, as most people don't have a good sense of what they'd like to do with AI yet.
I strongly suspect Apple has a better idea than you do, though.
My bad here. The M3 has all the features listed for the M4 including mesh shading. However, it's still true that they had to redo layout, and I actually know for a fact that's not all they did, because I know someone who worked on the M4 GPU (and who refused to tell me any interesting details, sigh). Did that work amount to more than taking advantage of FinFlex? No idea. Maybe we'll know once the M4 arrives in Macs.
All the other stuff I wrote stands.