Will M3 Ultra be fused M3 Max, or?

[theorist9]

Especially when, as a feature set, ALL the Ultra has to be is “the fastest Mac”. Doubling up is a cost effective way of getting to “the fastest Mac”.

IMO, that would only be the case if Apple operated in a vacuum. But, as I'm sure you know, it doesn't. Apple's Mac division competes with Intel, AMD, and NVIDIA. Apple itself recognizes this when it compares the Ultra to PC's in its own marketing and promotional materials (see screenshot).

Thus Apple doesn't just want the Ultra to be the fastest Mac, it wants it to be fast relative to its competitors (not necessarily the very fastest, but fast enough to be respectable relative to PC workstations).

 

[deconstruct60]

That's a good point that the transistors used specifically to interface with the bridge wouldn't be needed. Though I don't know how much space that would save.

It is not a ton of space on an individual die.

Apple Announces M1 Pro & M1 Max: Giant New Arm SoCs with All-Out Performance

www.anandtech.com

Back of the envelope. The M1 Max is 19 mm wide. Even about 1x19mm^2 is in the ~20mm^2 zone. 20/430 is ~5% . That isn't going to buy anywere close to a 800mm^2 to 680-700mm^2 shrinkage move.

If multiply that by 1M or 10M that adds up when aggregate that large. That is a huge area savings.

 

[deconstruct60]

IMO, that would only be the case if Apple operated in a vacuum. But, as I'm sure you know, it doesn't. Apple's Mac division competes with Intel, AMD, and NVIDIA. Apple itself recognizes this when it compares the Ultra to PC's in its own marketing and promotional materials (see screenshot).

Thus Apple doesn't just want the Ultra to be the fastest Mac, it wants it to be fast relative to its competitors (not necessarily the very fastest, but fast enough to be respectable relative to PC workstations).


View attachment 2318344

Apple is not trying to compete with the entirety of AMD, Intel, and Nvidia's line up at all. Intel and AMD have more CPU processors SKUs with > 16 cores than Apple has entries in their entire Mac line up.

Apple picks targeted subset of the market. They are not trying to sell everything to everybody. Apple dumped macOS Server. The notion that Apple is going 'whole hog' trying to make a HEDT "high core count war' CPU SoC is misguided. They aren't.

Apple's chart is for the mainstream, upper end, non ECC desktop pool of the relatively large PC workstation market. Apple is 'overachieving' in the chart becuase they don't have an even higher end option for workloads that scale on core count.


In terms of single thread drag racing the Ultra is effective against the PC in that is can deliver about same ST as the rest of the M-series line up of that generations. Not seeing either 'tail off' as increase core count or deeply flattening per/power curves as clock past what is reasonable return on throwing more power consumption at the problem ( flogging the design past where the fab process really wants to be. ) . Apple isn't engaged in selling maximum Turbo overclock speed with exotic cooler #42.

 

[Chuckeee]

Back of the envelope. The M1 Max is 19 mm wide. Even about 1x19mm^2 is in the ~20mm^2 zone. 20/430 is ~5% . That isn't going to buy anywere close to a 800mm^2 to 680-700mm^2 shrinkage move.

Just an aside

Does anyone yet know the dimensions of the M3 die? Apple not telling (specific numbers) but I was wondering if someone somewhere has broken up a new M3 an actual measured the darn thing.

 

[Allen_Wentz]

There is no evidence for two separate M3 Max dies, it's likely binned/partially disabled die configurations. And there is also no reason to assume that M3 Ultra won't use the same dual-chip packaging as the Ultra chips before it.

I agree, for the Studio boxes. IMO Mac Pro needs some special new architecture because where the MP is now is not enough.

 

[casperes1996]

Interestingly, it actually is viable (see my last post; I also thought it was too big to etch until I looked into it a bit further), but they won't do it b/c it's cost-prohibitive.

Thanks for your post, Theorist. It was an interesting read. When I said it wasn't viable I more meant that even if it were technically possible I don't think it would be viable in a wholistic sense, considering cost and the like as well, hadn't at all looked into things like reticle limit although I also anticipated it was approaching that. But I appreciate your elaboration on the feasibility

 

[theorist9]

Apple is not trying to compete with the entirety of AMD, Intel, and Nvidia's line up at all. Intel and AMD have more CPU processors SKUs with > 16 cores than Apple has entries in their entire Mac line up.

Apple picks targeted subset of the market. They are not trying to sell everything to everybody. Apple dumped macOS Server. The notion that Apple is going 'whole hog' trying to make a HEDT "high core count war' CPU SoC is misguided. They aren't.

It's a misrepresentation to characterize my post as "misguided" when you're just repeating what I said in different words...that Apple wants to be:

not necessarily the very fastest, but fast enough to be respectable relative to PC workstations

And the number of SKU's AMD and Intel have is irrelevant, as are their core counts. What matters is performance. Yeah, the AMD EPYC 9554 has 64 cores, and the M2 Ultra has only 24, but their GB6 MC scores are 20,172 and 21,323, respectivey. [There may be faster AMD's, but that's the AMD with the highest MC score on Primate's processor chart.] So Apple's CPU performance is respectable. And the fact that it gets crushed by a dual-EPYC workstation doesn't change that. Where Apple does have their work cut out for them is on the GPU side, which you didn't mention.

 

[flybass]

How much more are you going to pay? It is not like exchanging the same amount of die area for that substitution. An even more expensive package for an even smaller set of users is pretty likely a pricing death spiral.

In addition, primarily things will not work for a GUI operation system if it does not have any GPU cores to drive the GUI. All 'P cores' is not targeting macOS is any effective way at all.

E cores can be thrown at stuff like software RAID, network storage overhead, and other tasks on systems with even more 'background' overhead . ( open activity monitor and look at the number of processes on a typical macOS instance. there is lots more than the one app that might be running on the screen . ) The M3 Max 12-4 ratio ( 3:1) is actually higher than the M2 Max 8-4 ratio ( 2:1). So probably didn't really spend more die area budget on the E cores here at all.

There is substantive upside in putting 'scut work' stuff into the local L2 cache of the E cores that helps keep P core L2 caches 'cleaner'. Don't really want P cores on 'scut work' tasks.

Tossing out the E cores isn't going to save any substantive space (relative the other major core cluster types). Probably wouldn't even get you another display controller. So to go to all P cores would have to eat into something else's die area budget. The OS already (and a substantial number of apps) has a substantial amount of task delegation built into it so not like run into major problems trying to engage E cores when should not.




The M3 Max is better than M2 Max. And the Ultra is extremely likely going to get the same multiple over the M2 version. That is more maketspace relevant. It doesn't get Apple a 'Threadripper killer' SoC , but Apple Mac Stduio doesn't need a Threadripper killer SoC.

Thanks for the reply. Given background tasks can't be unique to macOS, do workstation intel and AMD processors have a similar fraction of e-cores? I'm asking mostly out of curiosity.

 

[theorist9]

Thanks for the reply. Given background tasks can't be unique to macOS, do workstation intel and AMD processors have a similar fraction of e-cores? I'm asking mostly out of curiosity.

As least for Intel, their E-cores are not comparable to Apple's. Intel's E-cores are much more powerful (and power-hungry) than Apple's. They're more liked lower-clocked P-cores. Different design philosophy. Also, Intel doesn't have any single ratio. It varies across its product line. As for the specific numbers, you can google those pretty easily and create your own table, if that's of interest to you.

Having said that, at least looking at Intel's Core series processors (which are used in workstations, as you can see from Boxx's website), Intel typically uses a higher ratio of E : P than Apple. My theory is that, because Intel's P-cores are so energy-demanding, Intel needed a way to offload a greater percentage of computing demand to the E-cores than is the case with AS. That would explain why Intel's E-cores are both more more numerous and more powerful than Apple's.

I've read Intel's next generation of Xeon's will feature E cores. I don't know if the Xeon's have E cores now.

 

[leman]

I agree, for the Studio boxes. IMO Mac Pro needs some special new architecture because where the MP is now is not enough.

The only way up for Apple right now is to go quad Max die.

 

[theorist9]

The only way up for Apple right now is to go quad Max die.

And TSMC recently announced a new process for bridging multiple dies, which Apple might make use of. But probably not before 2025—which, yes, is not right now.

 

[Chuckeee]

And TSMC recently announced a new process for bridging multiple dies, which Apple might make use of. But probably not before 2025—which, yes, is not right now.

Probably a much nearer term solution than chiplets.

 

[Allen_Wentz]

The only way up for Apple right now is to go quad Max die.

Apple might think outside the box. We will see. I am not a chip engineer, but it seems to me that folks must be thinking about layers in addition to thinking about sticking chips together at the edges like they do with Ultra or a double-Ultra. Heat removal is of course an issue, but that is what makes it all so interesting...

 

[Unregistered 4U]

IMO, that would only be the case if Apple operated in a vacuum. But, as I'm sure you know, it doesn't. Apple's Mac division competes with Intel, AMD, and NVIDIA. Apple itself recognizes this when it compares the Ultra to PC's in its own marketing and promotional materials (see screenshot).

You’ll notice the comparison is with a processor that wasn’t indicative of Intel’s peak at the time. The idea that they “compare well” in March of 2022 to a “Core i9-12900K and DDR5 memory” solution released in November of the prior year (they didn’t even bother to compare it against the more performant version released just that January) really puts a fine point on how much Apple is concerned with the competition. Apple knows that anyone getting an Ultra powered system is more concerned about it “running macOS faster than any other currently available option in that form factor” than they are about it performing like a middle of the road PC.

And all of this is to say, THAT is what’s behind the Ultra being “just more of the same” and not some completely different architecture. There’s really no option for folks that need macOS and peak performance. And, as the company making the fastest shipping macOS computers, whatever they put out will be the fastest (and they’ll find some middle of the road Intel processor to draw a curve against)!

 

[Apple Fan 2008]

Given the separate dies for the different M3's so far, what are the chances that the M3 Ultra is an entirely separate die as well?

Possibly, if they do an “extreme” version that would probably be 2 Ultras fused

 

[flybass]

You’ll notice the comparison is with a processor that wasn’t indicative of Intel’s peak at the time. The idea that they “compare well” in March of 2022 to a “Core i9-12900K and DDR5 memory” solution released in November of the prior year (they didn’t even bother to compare it against the more performant version released just that January) really puts a fine point on how much Apple is concerned with the competition. Apple knows that anyone getting an Ultra powered system is more concerned about it “running macOS faster than any other currently available option in that form factor” than they are about it performing like a middle of the road PC.

And all of this is to say, THAT is what’s behind the Ultra being “just more of the same” and not some completely different architecture. There’s really no option for folks that need macOS and peak performance. And, as the company making the fastest shipping macOS computers, whatever they put out will be the fastest (and they’ll find some middle of the road Intel processor to draw a curve against)!

Your argument about Apple's strategy makes sense. Having said that, personally I'd never consider buying a studio over a MacBook Pro given the current state of things.

 

[Sydde]

As least for Intel, their E-cores are not comparable to Apple's. Intel's E-cores are much more powerful (and power-hungry) than Apple's. They're more liked lower-clocked P-cores.

That is not correct. Intel e-cores are a lot smaller than p-cores, on the order of 20% the size, which is why there are more of them on a chip. P-cores have much more SIMD capability than e-cores, more L1, and only p-cores have HT (which is to make use of the extra space they take up, and which is ironically not the best way to use HT). They draw a lot of juice because the architecture is expensive to parse and run.

 

[leman]

Apple might think outside the box. We will see. I am not a chip engineer, but it seems to me that folks must be thinking about layers in addition to thinking about sticking chips together at the edges like they do with Ultra or a double-Ultra. Heat removal is of course an issue, but that is what makes it all so interesting...

Yes, and Apple had patents describing these kinds of solutions for years now. But it won’t arrive with M3 family.

 

[Allen_Wentz]

Your argument about Apple's strategy makes sense. Having said that, personally I'd never consider buying a studio over a MacBook Pro given the current state of things.

Why not? What does the current state of things have against the Studio? [Other than the obvious fact that there is no M3 Studio yet]

This from someone that bought an M2 MBP only because Apple delayed the M2 Studio (like they are delaying the M3 Studio now). I now have two MBPs, one old. At some point (2024-2025 maybe) I will probably replace the older MBP with a Studio. But maybe not if the 2016 MBP keeps chugging along OK for web use.

 

[flybass]

At least for me, the compute power of the ultra doesn’t compare favorably with compute power of max + portability option of a mbp.

 

[Allen_Wentz]

At least for me, the compute power of the ultra doesn’t compare favorably with compute power of max + portability option of a mbp.

I fully agree that most folks will benefit from having a laptop, and Apple made the first true desktop-replacement laptop in 2011 when Apple (exclusively for 2011) brought Thunderbolt i/o to the world. I have owned one ever since, currently an M2 MBP with 96 GB RAM.

 

[theorist9]

That is not correct. Intel e-cores are a lot smaller than p-cores, on the order of 20% the size, which is why there are more of them on a chip. P-cores have much more SIMD capability than e-cores, more L1, and only p-cores have HT (which is to make use of the extra space they take up, and which is ironically not the best way to use HT). They draw a lot of juice because the architecture is expensive to parse and run.

Nope. When the M1 came out, Anandtech estimated that an E-core offered about 20% the performance of a P-core. Since then their E-cores have advanced faster than their P-cores, but they're still not comparable to Intel's. Feel free to find your own numbers but, according to this, Intel's E-cores offer 54% of the performance of a P-core without hyperthreading, and 40% of the performance of a P-core with hyperthreading:

https://www.reddit.com/r/intel/comments/yv3uan
[The benefit from HT is of course going to be highly task-dependent, so with different tasks the disparities would be different.]

See also this review [ https://www.techpowerup.com/review/intel-core-i9-12900k-e-cores-only-performance/8.html#:~:text=We see the single Gracemont,3900X, not too far ahead. ], which said:

"We see the single Gracemont E-core achieve a score on-par with the Ryzen 3 3300X. What this means is that Gracemont has an IPC roughly comparable to AMD's "Zen 2" in math-intensive, low-bandwidth workloads. We see other Ryzen 3000 series chips, such as the 3600X and 3900X, not too far ahead."

All of this fully supports what I said:

As least for Intel, their E-cores are not comparable to Apple's. Intel's E-cores are much more powerful (and power-hungry) than Apple's. They're more liked lower-clocked P-cores. Different design philosophy.

Note that I didn't get into technical details, since the point was to discuss bottom-line performance and power consumption. And an E-core that provides about half the power of a P-core is indeed more like a lower-clocked P-core, in terms of its processing power and the kind of tasks it would be used for, than is the case with Apple's E-cores. I.e., while Apple's E-cores appear to be designed for background tasks, Intel's are sufficiently powerful to help the P-cores with foreground tasks.

 

[altaic]

And TSMC recently announced a new process for bridging multiple dies

Link?

 

[leman]

@Sydde @theorist9 Sydde is correct. Intel’s E-cores are much smaller than P-cores (still quite large in comparison to AMD and Apple). Intel’s E-core are about area-efficient compute first and foremost. But theorist9 is also right that they are relatively fast. All in all, I don’t think that there is a disagreement here.

Here is a die size analysis with core areas for both Alder Lake and the upcoming Meteor Lake: https://www.semianalysis.com/p/meteor-lake-die-shot-and-architecture

 

[Sydde]

It would be interesting if Apple made an ultra with all performance cores.

When I look at GB, I see the MC score of a 13900 beating a M3 Max by about 3%. The i9 is 16E+8P, meaning it can run 32 threads at once, while the Max is 4E+12P, giving it no more than half the thread capacity, yet it is millimeters behind in the benchmark. Add one core cluster (four more P) and it will be faster than anything Intel has to offer, without having to go all-P.