MP All Models 40 core ASI with PCI-E slot is around the corner!!

[deconstruct60]

Only possible if M3 does not use 3nm node. Every bit of 3nm manufacturing capacity will be reserved for iPhone 15 this year, you can be sure of that.

Digitimes reported that TSMC N3 production capacity was at 45K wafers per month now and standing at 50% utilization. There is very large under utilization of TSMC capacity at the moment. Every last drop of N3 capacity is NOT needed for iPhone 15 , because it is not even being used now. If Apple was that desperate they would be using it.

One huge issue here is that it is NOT the iPhone 15 that needs N3. The regular iPhone (and Plus) are extremely likely going to use the A16 chip; not A17. So it is ONLY the Pro models that need the newer chip. That saves 10's of millions of SoC 'need' off the books.

They could have just stockpiled those in January . Similarly for the plain M3 ( if did some in Feb-April before the A17 really got going then would have enough to overlap.).
At about 10K wafers per month Apple could probably all the M3 Ultras they'd need to handle the initial demand bubble for the Mac Pro .
They would only need about 8-10K wafers to do a 100K M3 Ultras.

https://forums.macrumors.com/threads/apple-orders-entire-supply-of-tsmcs-3nm-chips-for-iphone-15-pro-and-m3-macs.2381637/post-31996254

That is more than they would probably ever sell in a whole year if keep the some > $6K price points. They could have just stockpiled those in January . Similarly for the plain M3 ( if did some in Feb-April before the A17 really got going then would have enough to overlap.).

To do 10M A15 sized SoC per month is only around 20-23K wafers per month. If TSMC N3-family capacity is > 45K/month by June that shouldn't be a problem. The only "problem" the iPhone SoC has with N3 capacity is the timing of the initial demand bubble. The steady state iPhone pace isn't a problem. The 'short term' initial bubble demand is the problem. The 'easy' solution to that is to just start a month or two earlier than normal on production.

A M2 sized die ( not any good reason the N3 based M3 shouldn't be a bit smaller, but stick with larger size) is about a 27-30K wafer per month issue to do a whole 10M (and only 5K wafers per month to do 2M/month). So TSMC is up to 45K could handle those if not spiking at the same time. [ move the M3 demand bubble well away from the A17 demand bubble ] .


TSMC has capacity to handle the load if spread the fabrication out over the whole year. What they don't have is the capacity to do "everything" all at the last minute. Which really doesn't work for N3 anyway since its "bake"/fabrication times are relatively longer. It is not a "do 30 million all at the last possible minute" kind of process.

The problem that Apple has is really not the iPhone so much as other players jumping into N3E in the second halve of the year. That's is what is going to crank the N3-family fab capacity at 100%; the other folks.

 

[Boil]

Would be sweet to see Apple introduce the M3 family with the top-end variants, the M3 Ultra & M3 Extreme (or whatever Apple decides to call them), in their top-end headless desktop, the ASi Mac Pro...

 

[deconstruct60]

Whatever even a factual mind like yours should admit too many coincidences not just on rumours which seem incoherent as no m2 extreme, instead two M2-Ultra (still 4 m2max but no m2 extreme),

The "two M2 Ultra" rumors don't make any sense. UlraFusion only works on extremely short distances. (talking less than an 1 inch ( or ~3 cm). It is a whole different ballgame to try to lash together multiple packages that are going effectively multiple inches apart ( the RAM modules are extremely likely to get in the way). The basic design premies of UltraFusion basically fall about at that part. So when the "two Ultras" hand waves at a 'modified UltraFusion" connector between the two ... it is just that ... highly likely mainly hand waving.


Just pretty dubious that Apple is going to invent another connector that is much lower Perf/Watt as well as substantively lower bandwidth when they have a baseline connector tech that is better at both.

What Apple has is a bad chiplet design in the Max die. To go to 'better than Ulra" they just need something far better designed... not pulling out a bigger hammer to pound a round peg into a square hole.

M2-max includes double sized buses indeed new UltraFusion foresee double capacity.

I have seen some comments about better internal bandwidth inside the M2 Max but that only puts even higher pressure on keeping the UltraFusion connector as short as possible.

Note on the M2 Max though that the aggregate Memory bandwidth did NOT get higher. The faster internal bus is mainly so that the SoC can better walk and chew gum at the time (e.g., do a more Video en/decode concurrently with more GPU cores grabbing/saving data while also running the P core AMX as full speed. ) . The internal pressure to feed/update/share caches is higher .... so the internal bus bandwidght got better. That makes going 'off die' harder ; not easier.

Which is why some contrived, Rube Goldberg , off die set up might have failed ( e.g, too many more additional NUMA problems ) .

As for the 'new' UltraFusion ... sure wasn't in the die pictures so far. If folks have shots of the UltraFusion still being there I haven't seen it yet. Even if it is there , then it is somewhat likely some late breaking kludge ( if on a M1 , M3 , M5 sequence) than something to crow about (or even scales reasonably well).

Still pending on what seems an eGPU or compute module what it would be. But this trashcan cheese grater hybrid perfectly fits everything apple wants from us, and admit the trashcan is so divisive something Apple marketing enjoy.

With the iOS tie-in haven't seeem much even remotely creditable evidence that this is even eGPU/GPU related. Just as likely part of some mutiple SoC set up in the headset as anything "Mac" related.

 

[Boil]

As for the 'new' UltraFusion ... sure wasn't in the die pictures so far. If folks have shots of the UltraFusion still being there I haven't seen it yet. Even if it is there , then it is somewhat likely some late breaking kludge ( if on a M1 , M3 , M5 sequence) than something to crow about (or even scales reasonably well).

@Mago claims to have x-ray proof of a double UF connection, but when asked for the imagery they said "my personal work is private", which is a bit of a cop-out; an x-ray of a commercially available laptop should not have any real need to be kept a secret...?

 

[treehuggerpro]

Dear Gawd (Apple), please make it stop!!!

Dear Gawd (Apple) I know, I know, I’ve sinned against you (by reading this thread). I was desperate, I was weak, I had my doubts, I was impatient and ignorant, I wanted a sign, just an itty-bitty sign.

Dear Gawd (Apple) I was tempted, the dark was so so overwhelming, but I hereby do repent, no more fantasy, no more la-la land, no more sneaky peaky around the corner!! (Apologies @Amethyst)

Dear Gawd (Apple), I surrender myself unto (and until) Thine Truth and (Ultra Raytraced) Light be known here on Earth (as it is in Cupertino), Forever and Ever, Amen!

 

[PineappleCake]

they can add some m.2 slots at the very least?
Or apple can spin an 4 slot apple storage raid0 setup as the fastest storage ever in an workstation.
But who would buy that at apples prices and have an raid 0 setup that needs high cost parts and DFU mode to restore (if apple lets you buy just say card 3 and not force you to buy an full set of cards just to repair 1)

Apple has never used M.2 slots

 

[leman]

Chinese forum speculate the New Mac Pro is in trashcan form factor, with one or two M2-Ultra linked by an New UltraFusion to UltraFusion fabric, if not enough bad news, only storage is discreet, 6x TB4 and 4 USB 3.1 type A like the Old trashcan "can't innovate my ass anymore" , and comment it to be released at WWDC keynote and not at late April.

What's the background of this speculation? Apple did recently (like last week) publish a continuation of the trashcan patent... but all of this does sound very far fetched.

So when the "two Ultras" hand waves at a 'modified UltraFusion" connector between the two ... it is just that ... highly likely mainly hand waving.


Just pretty dubious that Apple is going to invent another connector that is much lower Perf/Watt as well as substantively lower bandwidth when they have a baseline connector tech that is better at both.

I think there is one interpretation of this idea that could make at least some degree of sense, and that would be an Infinity Fabrik Link-style connector. Lower bandwidth than a die-to-die connection, sure, but sufficient to provide coherency between the CPU clusters, while the GPU could be exposed as a NUMA multi-device configuration, which is explicitly supported by Metal (just like the AMD duo cards).

What Apple has is a bad chiplet design in the Max die. To go to 'better than Ulra" they just need something far better designed... not pulling out a bigger hammer to pound a round peg into a square hole.

Why bad? What would they need to improve in your opinion?

I have seen some comments about better internal bandwidth inside the M2 Max but that only puts even higher pressure on keeping the UltraFusion connector as short as possible.

Going by the patent timeline it is likely that M2 Pro/Max use a new interconnect architecture, with separate networks specialised for GPU-to-GPU and CPU-toCPU communication. Combined with large caches, this is what likely gives the M2 family its significantly improved performance in Blender.

@Mago claims to have x-ray proof of a double UF connection, but when asked for the imagery they said "my personal work is private", which is a bit of a cop-out; an x-ray of a commercially available laptop should not have any real need to be kept a secret...?

Mago claims many things, most of them fairly outlandish... he always makes for an entertaining read though

 

[ZombiePhysicist]

Please let this be true. If it's another trash can, I'm buying the second it's on sale. I looove my 2013 trash can and the ONLY reason I would avoid the nMP is if it's a giant tower. That design is heavy, unnecessary, and antiqued.

wow if only there was a totally limited un upgradable closed system for you to buy… A stud machine with good I/o. A stud w/ io. Are you getting it?

 

[Mago]

What's the background of this speculation? Apple did recently (like last week) publish a continuation of the trashcan patent... but all of this does sound very far fetched.


Going by the patent timeline it is likely that M2 Pro/Max use a new interconnect architecture, with separate networks specialised for GPU-to-GPU and CPU-toCPU communication. Combined with large caches, this is what likely gives the M2 family its significantly improved performance in Blender.


Mago claims many things, most of them fairly outlandish... he always makes for an entertaining read though

First and foremost: it's just an speculation, neither an April 1 joke, the ASi Mac Pro perfectly maybe a cube hexagon or classic tower design as 7,1.

But, while we May have long forgot the Trashcan, actually it still fits Apple product philosophy, Apple allegedly it never got updated due thermal restrictions which cornered it (i can confirm there where few 2nd gen trashcan prototypes that never managed beyond test even with copper thermal core).

But ASi put it again on the cards, these patents while (as usual neither anticipate an actual product) confirm Apple never trashed the trashcan form factor, even evolved it, trashcan haters: it doesn't confirm an new trashcan-like Mac Pro, but meaningful confirm the form factor is not dead.

We agree on this (at fact level):
We have an new Mac pro coming soon.
Gurman told m2-extremme (4 m2 max on single interpose) not happening (?)
A Mac Pro without quad m2-max , has no sense.
GPU Compute power while it may come from AMD (and is known apple is working with AMD on some product) but it doesn't rules it will replace ASi iGpu but as compute accelerators just like trashcan 2nd GPU (oops I mention the trashcan).
There are apple new patents for flexible substrate high speed interconnect, while unlikely to replace UltraFusion, maybe very good for an surrogate compute board like was the 2nd trashcan GPU, don't forget Xeon 2p motherboard interconnect both CPU with an bus barely faster than 2x the faster PCIe x16, so that flexible substrate interconnect perfectly may work as it's similar on the classic Trashcan (cTrashcsn now onwards), not as perfect as an true 4p UltraFusion but good enough.

What we speculate:

ASi form factor, ISA expansion, 4p much needed solution, discreet RAM.

Don't forget Apple invented also: COURAGE, and we know what it means.

ASi form factor, while asiTrashcan not what performance purists would recipe, it fits Apple design narrative, further it makes sense for all incongruent rumours on 2p/4p, discreet ram, and while it's GPU power (specially overclocked) just matches trashcan single duo GPU, it's good enough for 95% Mac Pro users the remaining 5% maybe covered by slave-mode Macs as "compute device" thru thunderbolt 4 (demoed past WWDC with redshift and few Mac studio as surrogate compute device).

Maybe this sudden trashcan resurrection is product of ASi UltraFusion difficulty to scalate to 4p, or was something planned from start.

Do ASi Mac Pro include PCIe slots (ISA expansion): i think it should, there is no better way to install a compute accelerator or PCIe storage or i/o devices, but remember apple's courage... They did in the past (cTrashcsn) nothing prevents it being it's future (ASi Trashcan).

Discreet RAM support: READ: unified memory not the same as soldered memory, M2 max is perfectly capable to interface discreet DDR5 ram, a Mac Pro with soldered ram has no sense period, but it doesn't means you will be able to profit from RAM discreet upgradeability: Even with STD CXL2-DIMM apple perfectly may lock the trashcan RAM modules to only apple approved ones, its an CXL2 inbuilt feature apple unapologetically may enforce.

 

[leman]

A Mac Pro without quad m2-max , has no sense.

Just a quick comment on this: I think it does make sense, but only if they can clock it higher. There was this rumor floating about M2 Ultra running at 4.2Ghz, and if they can do that, then probably running the GPU at 1.6Ghz is also not out of the question. This hypothetical configuration would deliver around 2300/35000 on CB5 for the CPU and 30+ TFLOPS for the GPU (or around 270-300k CB6 compute score). That would be sufficient to compete with the new Sapphire Rapids-WS (except the top w9-3495X) on the CPU front and the 7900xtx on the GPU front (while AMD has more FLOPS on paper, it's dual-op capabilities are limited in practice, so the advertised numbers are mostly BS). But if they retain the clocks of the current M2 family, the M2 Ultra would be not very interesting indeed.

As to the rest, I think all you wrote is very reasonable.

 

[Mago]

There was this rumor floating about M2 Ultra running at 4.2Ghz

I buy it will be overclocked at some range but I don't think apple to push it beyond a TDP where it would require liquid or sophisticated cooling, i believe it would be below 20% matching clock delta among MBP' LPDDR5 and server grade LPDDR5.

Further ASi GPU performance not clock linear.

I think an overclocked quad m2-max solution to top close but below 70 TFlops.

And no, the ASi Trashcan neither will be anything good at gaming, would be excellent at machine learning, graphics rendering video processing, buy at gaming Even the quad m2-max model Will be barely better than the studio.

 

[leman]

I buy it will be overclocked at some range but I don't think apple to push it beyond a TDP where it would require liquid or sophisticated cooling, i believe it would be below 20% matching clock delta among MBP' LPDDR5 and server grade LPDDR5.

They should have enough TDP reserves that pushing the clock by 20% would still keep them in a good spot. I mean, even if we assume something ridiculous like a 3x increase in power consumption, we are still looking at ~200W CPU and ~300W GPU, easily manageable with a single large heatsink and a Mac Pro chassis. More likely though they can do at at 2x cost of power (or less), which would put them at the much more modest ~300-400W — a TDP of a large gaming GPU.

Further ASi GPU performance not clock linear

Why wouldn't it be? ALU performance scales linearly with clock and Apple GPUs have plenty of RAM bandwidth in reserve (I mean, Nvidia ships 16384 ALUs with 2Ghz+ clock and only has a 1TB/s RAM interface).

 

[avro707]

wow if only there was a totally limited un upgradable closed system for you to buy… A stud machine with good I/o. A stud w/ io. Are you getting it?

Maybe it’s a double height trashcan?

Wasn’t there a patent drawing for something like that?

 

[Mago]

we are still looking at ~200W CPU and ~300W GPU, easily manageable with a single large heatsink and a Mac Pro chassis

Gamer mindset, but not manufacturer mindset is to provide extensive warranty on machines often use to work on a single project 24x7 for weeks at full load, indeed whatever the overclock likely will be conservative.

Why wouldn't it be? ALU performance scales linearly with clock and Apple GPUs have plenty of RAM bandwidth in reserve

GPU/clock performance from m1 to m2 didn't scaled linear, beyond "Max-Q" point it degrades bot on yield and efficiency.

 

[goMac]

I’m just over here remembering how all the speculation this forum had about the 7,1 was completely wrong.

 

[PineappleCake]

Gamer mindset, but not manufacturer mindset is to provide extensive warranty on machines often use to work on a single project 24x7 for weeks at full load, indeed whatever the overclock likely will be conservative.

What?? AMD workstation chips run at 200+ watts and you can get them from Lenovo.

Apple can do this otherwise Zen 4 threadripper will destroy any Mac Pro CPU that is around 32 cores.

 

[leman]

Gamer mindset, but not manufacturer mindset is to provide extensive warranty on machines often use to work on a single project 24x7 for weeks at full load, indeed whatever the overclock likely will be conservative.

Well, that's the question, isn't it? Can they ship higher clocked configurations without compromising stability and long term reliability? We simply don't know the answer to that, but if they can do a 20-30% overclock, then M2 Ultra would work just fine as a workstation chip. The TDP is only of secondary value here as it's not a measure of system reliability. We have professional systems running at super high wattages. Hell, Apple runs the Xeons in the 2019 chassis at over 300 watts sustained and they provide 500 watts of power per MPX module.

GPU/clock performance from m1 to m2 didn't scaled linear, beyond "Max-Q" point it degrades bot on yield and efficiency.

Not sure what you are basing this on? GB6 shows roughly 12% better performance per core, which is exactly the clock difference. In gaming benchmarks (I looked at Wild Life Extreme) the improvement is closer to 15% per core. And when you look at Blender rendering, M2 Pro/Max are much faster than M1 Pro/Max (we are talking about ~60% improvement per core). But M2 series bring more than just a clock improvement, the M2 Pro/Max also have more L2 cache and likely large internal improvements to the GPU data bus and shader dispatch.

At any rate, 20-30% higher clock on these devices should see 20-30% higher computer performance across the board unless the workload is bandwidth limited. It's exactly the same as with third-party GPUs.

 

[leman]

Apple can do this otherwise Zen 4 threadripper will destroy any Mac Pro CPU that is around 32 cores.

An M2 Ultra with 20-30% higher CPU clock than M2 Pro/Max should perform similarly to a 32-core EPYC Zen4, which is more than sufficient for Apple. They are not going to compete against upcoming 64-core or 96-core Threadrippers any time soon, but that's also not the market they are looking at. At any rate, such an M2 Ultra should have multicore CPU performance similar to a 5995WX, with substantially better single-core performance.

 

[majus]

April 2 Gurman on the Mac Pro:
"Beyond the headset, I’m told that multiple new Macs are in the works for launch around the middle of the year. The lineup could include fresh MacBook Air models and a Mac Pro with in-house Apple chips. The latest iOS and macOS beta versions will build in support for the new machines. We’ll also get a look at iOS 17, iPadOS 17, watchOS 10, tvOS 17 and macOS 14."

So now it's maybe May.

 

[enc0re]

WWDC June 5, https://forums.macrumors.com/thread...-2023-event-taking-place-june-5-to-9.2385254/

24h earlier than I expected.

Next announcement ASi Mac Pro, sometime in April likely before AFTER Easter April 11 or 18.

---

Edit: 🤯🤭

According to Mago, we only have 9 or 16 days left to the launch of the Mac Pro. So we’ll know very soon…

 

[Mago]

The TDP is only of secondary value here as it's not a measure of system reliability. We have professional systems running at super high wattages. Hell, Apple runs the Xeons in the 2019

People here hated the trashcan among other reasons by its poor thermal flexibility, not up for long renderings, apple won't risk again on delivering a product which may overheat.

 

[Mago]

According to Mago, we only have 9 or 16 days left to the launch of the Mac Pro. So we’ll know very soon…

Yes we should see a new Mac pro those days, unless speculation on WWDC23' banner is true an that shadow is the ASi Trashcan

 

[leman]

People here hated the trashcan among other reasons by its poor thermal flexibility, not up for long renderings, apple won't risk again on delivering a product which may overheat.

The 2019 chassis has plenty of cooling potential, so they would need to be shipping some serous firepower for it to risk overheating.

 

[goMac]

I don't think we'll see the Mac Pro until WWDC.

 

[ZombiePhysicist]

It seems weird to me that they would release a Mac Pro before WWDC since they just announced the show date. I guess crazier things have happened.

I guess if the machine really sucked a pr release could make sense not to draw attention to how much it sucks, maybe? But it otherwise seems ideal for WWDC release, where they could boast about it a bit.