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
P-cluster AMX unit.
So you can't tell an AMX unit from an E-cluster....
Not chopped down die, chopped down die design.
BTW, only the M3 Max GPU has this extra cache. Each quarter of the GPU has one. They seem to be related to the 4 SLC blocks.
Attachments
Last edited:
For the Max, the big question is if various uncore aspects (NoC, etc.) are a significant win over M2. If so it'll be a bigger win that you're thinking, probably. The 30-GPU-core version may or may not be a significant improvement over the M2 Max (aside from RT which is either huge or nothing, depending on use case); the "missing" bandwidth may not have much of an impact.
The M3 Pro is a bit of an open question. Yes, the P/E distribution and memory bus are potential negatives, but I think we need to see it in action before we decide. There are likely corner cases where the bandwidth is an issue, but very few - 150GBps is still quite a lot. (That's assuming they haven't done anything super-interesting like full memory compression, which would change the story completely.) 6P+6E will always be less good than 8P+4E of the same core types, but these are not the same as the M2, so I don't think we'll have a good sense of how they compare until benchmarks come out. My bet is that there will be no meaningful set of cases where the M2 is better, but facts > bets.
For the M3, there may be small gains from the uncore along with the advantages each core brings. But yes, that was always going to be an incremental upgrade on the CPU side. RT is pretty huge, where it's used, as are a few other things like AV1 decode, and meaningless otherwise.
But for all of these, even after we have benchmarks, we *still* won't know something important. How much clock headroom do these chips have? I suspect we will see these very same M3s run faster in desktops (minis and studios). And I think the 16" will run a little faster/hotter than the 14", though probably not by all that much.
Go ahead, throw your jabs in.
I posted my reasons why I thought it might not be A17 Pro here: https://forums.macrumors.com/threads/are-we-sure-m3-is-based-on-a17-pro.2409293/
It's just my speculation. No one can say for sure right now that it's based on A17 Pro. Perhaps it's modified. Perhaps parts of it was ported over. Perhaps Apple cut some things out of the M3 that was in A17 Pro.
Feel free to argue against my points instead of throwing these cheap shots in. @leman already explained why he did it and I accepted it. You though? Who are you? What exactly do you contribute to this forum?
I like your optimism but I think something like full memory compression would have received a call out in the presentation and the lack of performance data on M3 Pro is making me sceptical.
I have ordered a 14 inch M3 Max because M3 Pro looked a bit gimped and I think that's probably what Apple wants!
I agree, I think it's low probability. But there are other things they could be doing to improve effective bandwidth or reduce latency that could make a meaningful difference. We'll know when the benchmarks hit.
One more thought about these chips in desktops: It's possible that they were designed to support LPDDR5X as well as LPDDR5. If so we might see that in those Macs. Like full memory compression I don't think it's all that likely, but it's far from impossible.
Edit: One more thing I forgot to mention - there was no talk about PCIe implementations (4 or 5?), nor about SSD performance. This is ripe for improvement - Macs have had (mostly) strong performance on their SSDs for a while now, but they're starting to get slaughtered by NVMe drives on PCIe 5. It would be nice if Apple did something about that too. We won't see it in the laptops or they'd have said something about it, and that's no surprise - those things suck power! But in the Mini or Studio, it could easily be done.
Edit: One more thing I forgot to mention - there was no talk about PCIe implementations (4 or 5?), nor about SSD performance. This is ripe for improvement - Macs have had (mostly) strong performance on their SSDs for a while now, but they're starting to get slaughtered by NVMe drives on PCIe 5. It would be nice if Apple did something about that too. We won't see it in the laptops or they'd have said something about it, and that's no surprise - those things suck power! But in the Mini or Studio, it could easily be done.
Last edited:
It’s different. I wish I had a decent die shot of the A16 to compare, but the only one I’ve seen floating around is crap quality. I suspect the M3* NPUs are a new design.
How sure are you? The quoted performance numbers are probably different than the A17 because they're quoting different sizes (INT8 on A17 IIRC). That's not what they used for the M2 announcement, so they may have opted for consistency with that.
That's a super interesting observation! So the ratio is now one AMX block per six cores? And it looks significantly larger in relation to the P-core than on M2, right? I am curious what this means for the cache hierarchy... having six cores sharing L2 cache could be nice for complex multithreaded workloads. Not to mention that this likely means more L2 cache for single-core operation?
But is that the same as A17? Here is the best shot I have seen:
Ah, sorry, I misunderstood what you were saying. So weird... I don't really understand why they wouldn't reuse the block they already had?
No idea. FWIW, the M3 NPU looks a bit larger than the A17Pro NPU, indicating that the design difference wasn’t to save die space.
Just speculating, but do we know those die shot images are actually real? Or at they just CG representations of the architecture?
Fascinating and unexpected. My guess would be that the A17P was optimizing for INT8 (and INT4?) while they did something in the M3 to optimize handling 16-bit values.
BTW, this guy’s floor plans look pretty much correct to me (he’s just missing a few things):
Last edited:
From what I understand Apple posts real enough images. This might be CG but the layout is accurate.
Apple has tampered with their die shot in the past to hide the existence of the M1 Ultra, so while it’s possible that they’re not exact, I’d be very surprised if they were outright fabricated.
No they are not E cores. I mark additionally with the E clusters in green and their AMX in white. The M3 Pro one is clearly larger.
It is quite interesting that the NPU has different floor plan in three socs and some of them are not even in square now:
Thanks, sorry about that. They shift in position relative to the P cluster on each die so I had trouble picking those out. And the Max is I/O on top and the other two are I/O on bottom (and GPUs vice versa).
We will see. Quoting Apple, both M2 Pro and M3 Pro has a 20% generational improvement on CPU over M1 Pro
I, for one, am on your side.
There's way too much noise being generated by people based on god-knows what, but certainly not any technical competence.