M4 Mac Announcements

[dada_dave]

For easier comparison, I've updated my Mx table. Now I wonder if we are ever going to see an M3 Ultra...

We can pretty definitively say no at this point.

 

[theorist9]

For easier comparison, I've updated my Mx table. Now I wonder if we are ever going to see an M3 Ultra...
EDIT: Fixed a typo.

Nice. You can add the M4 Max info based on what Apple just released.

 

[theorist9]

Will we be able to tell from the M4 die shots whether the M4 Ultra will be a 2x M4 Max? I.e., can you say whether it contains an interface for the connection bridge? I vaguely recall some discuss of this for the M3 Max.

 

[dada_dave]

Will we be able to tell from the M4 die shots whether the M4 Ultra will be a 2x M4 Max? I.e., can you say whether it contains an interface for the connection bridge? I vaguely recall some discuss of this for the M3 Max.

Should be able to, yes.

 

[theorist9]

Should be able to, yes.

What was the story with the M3 Max?

 

[dada_dave]

What was the story with the M3 Max?

No interconnect was visible on 3rd party die shots unlike for the M2 Max. Of course people still thought we might get an M3 Ultra, which obviously never happened, so that spawned (including from me) all sorts of theories about what Apple might have planned.

 

[NotEntirelyConfused]

Yes, that and the 273GBps memory are better than even optimistic predictions.

And what about TB5??? I would have bet against that. The chip is a tour de force.

Now, as for the Max, I see four options:
1) Just more GPU, same mask as Pro, uncut. I hope not, but plausible. Same idea as M1.
2) Just more GPU, same mask... but, the sixth P core in each cluster is enabled. So, 12P + 4E.
3) Another P cluster, more GPU. 15 or 18 P, 4 or 6 E. 30-40 GPU cores. Still the same mask.
4) Different mask, all bets are off.

Well now I'm grumpy. I really wanted more from the Max.

We still don't know which door Apple picked. It could be 2, though that seems somewhat wasteful. And it could be 4. No way to know until we get die shots.

This does tell us something really interesting though. ISTM this shows us Apple trying to triangulate their optimal market positioning for the Pro. They tried 150% baseline in the M3 generation. Now they're trying ~200% in the M4 gen.

 

[theorist9]

No interconnect was visible on 3rd party die shots unlike for the M2 Max. Of course people still thought we might get an M3 Ultra, which obviously never happened, so that spawned (including from me) all sorts of theories about what Apple might have planned.

Ah, right--that was the "Hidra" chip, yes? Is there any consensus among the rumors about whether the 2025 Studio will use the M4 Max, or something newer?

 

[KingOfPain]

Nice. You can add the M4 Max info based on what Apple just released.

I didn't even think to check if there was another release...
EDIT: Corrected M4 Max RAM sizes and removed the elusive M3 Ultra.

 

[leman]

This does tell us something really interesting though. ISTM this shows us Apple trying to triangulate their optimal market positioning for the Pro. They tried 150% baseline in the M3 generation. Now they're trying ~200% in the M4 gen.

It looks like they went back to M1/2 family system. M3 so far is an outlier with a weaker Pro. I guess the customers didn’t like it that much?

 

[theorist9]

I didn't even think to check if there was another release...

View attachment 32361

On the M4 Max, it's 36 GB RAM on the low-end chip, and 48/64/128 on the high-end chip.
And you might want to add an asterisk that the 64 GB RAM option on the M4 Pro is available on the Mini only. [Unless there are other instances like this, in which case you'd want to list all of them for consistency, which would be too much work.]

 

[dada_dave]

Ah, right--that was the "Hidra" chip, yes? Is there any consensus among the rumors about whether the 2025 Studio will use the M4 Max, or something newer?

Actually this was before the "official" (Gurman) rumors of a Hidra desktop die for the M4 generation. This was the M3 and people were positing a desktop die or a brand new packaging solution and of course in the end it was that the M3 Ultra simply didn't exist! However, the Hidra rumor is still very much in play for the M4/M5. As for the upcoming Studio, it's unclear because we don't know if the Hidra die is only replacing the high end or to be used throughout. I *think* Gurman's last missive was that it would be like a Mac Pro only chip and the Studio would still be Max/Ultra. But that's far enough out I wouldn't trust his predictions 100% - he's saying the Studio won't come until Summer and the Mac Pro by the end of 2025. So we potentially have a long wait to find out. Again, his timing could be off.

It looks like they went back to M1/2 family system. M3 so far is an outlier with a weaker Pro. I guess the customers didn’t like it that much?

They also made more dies this time around. Previous generation was the A17Pro, M3, M3 Pro, and M3 Max. Now it's the A18, A18 Pro, M4, M4 Pro, M4 Max, and possibly Hidra (and of course possibly the Ultra which is not really separate die but still engineering resources will have to go into the interconnect if it exists). It may be they just wanted to simplify their efforts for the Pro/Max.

 

[theorist9]

It looks like they went back to M1/2 family system. M3 so far is an outlier with a weaker Pro. I guess the customers didn’t like it that much?

It's interesting that they both strengthened the Pro and somewhat weakened the lower-end Max by offering it with 36 GB only. You can actually get more RAM with the lower-end Pro chip (48 GB on MPB, 64 GB on Mini) than you can with the lower-end Max chip! Though the RAM bandwidth remains higher on the Max (as it needs to be, since it has more GPU cores than the Pro).

I think it's a good thing that they've increased the max RAM on the Pro to 64 GB, for those who need more RAM but don't need a lot of cores. And they probably eliminated the 96 GB option on the lower-end Max because few were ordering it--it was too much of a jump. Don't know why the don't offer, say, 48 or 64 GB with the lower-end Max. Maybe the number and size of modules doesn't work.

 

[Jimmyjames]

I have to disagree. I think putting more CPU cores in the Max would just make the thermal situation uncomfortable (by Apple's standards). The die will be interesting to see if there are extra disabled cores in the Pro die or if the Pro die in addition to being chopped also has a slightly different cluster configuration. I believe @mr_roboto's point was that it wouldn't be too hard for Apple to do that. I actually feel it's the Pro CPU that is awkward here, again I would have preferred them trade a couple of CPU cores for more GPU cores.

The progression is a little awkward, especially for the CPU:

P+E CPU/GPU
M4 4+6/10 -> M4 Pro 10+4/20 -> M4 Max 12+4/40

I just feel like in the Pro price tier the CPU power is overkill while the GPU power is a little underwhelming. I mean the GPU is okay and at 2 extra cores plus a speed bump should be a nice boost over the M3 Pro, but I feel like it could be pushed a little more. But that may be a function of the M4 Pro being a chopped Max.

I agree. It seems that Apple is going for their old formula of "really fast CPUs and kind of meh GPU". I understand where this comes from, but it is a bit disappointing for those who need more GPU performance. Curious to see how Cyberpunk runs on the Pro version.

I’m curious. When you talk about meh gpu, is this applicable to just the Pro, or do you both feel the standard and Max are underwhelming gpus as well. While I am happy with the cpu situation mostly, and I think the underlying gpu tech is really good. I can’t deny I’d like a little more horsepower.

 

[dada_dave]

I’m curious. When you take about meh gpu, is this applicable to just the Pro, or do you both feel the standard and Max are underwhelming gpus as well. While I am happy with the cpu situation mostly, and I think the underlying gpu tech is really good. I can’t deny I’d like a little more horsepower.

I'm always going to say more is better but ... I think the base M4 GPU should be good given the competition - okay maybe in the 14" Pro it's a little weak for the price point but with 16GB of RAM, honestly it should be decent. The M4 Pro GPU I just feel like is about 20% lower than where it could be relative to say the 4050 and 4060 mobile. The M4 CPU could probably lose 20% and still be an amazing CPU for the price point and thermals. The resulting die would slightly bigger than what we have now, but not a lot. More problematic would be if the Pro SOC is a chop in which case you can't just rejigger everything like this at will (although if it is a chop and there are two fused off P-cores in the Pro SOC then a hypothetical custom Pro SOC as I describe it would actually be slightly smaller). The M4 Max will probably be a fine GPU - probably very equivalent to, actually probably better than, a 4070 mobile with more RAM, better bandwidth, and better thermals. Of course the 4080 mobile exists (oh and the 5000 GPUs will be released soon) and Apple might be a bit pricy at the Max level, but the combination of device characteristics + SOC (especially CPU), it's pretty damn powerful. So overall, I'd personally describe the Max GPU in the laptop as more whelming than underwhelming. But definitely underwhelming in the Studio desktop.

 

[theorist9]

Based on the bandwidth figures, I'm thinking:

M4: 64 bits/module x 2 modules x (7500 x 10^6)/s x 1 byte/(8 bits) x 1 GB/(10^9 bytes) = 120 GB/s
Memory options: 16 GB => 2 x 8 GB modules; 24 GB => 2 x 12 GB modules; 32 GB => 2 x 16 GB modules

M4 Pro: 64 bits/module x 4 modules x (8533 x 10^6)/s x 1 byte/(8 bits) x 1 GB/(10^9 bytes) = 273.056 GB/s
Memory options: 24 GB => 4 x 6 GB modules; 48 GB => 4 x 12 GB modules; 64 GB => 4 x 16 GB modules

low-end M4 Max: 128 bits/module x 3 modules x (8533 x 10^6)/s x 1 byte/(8 bits) x 1 GB/(10^9 bytes) = 409.584 GB/s
Memory options: 36 GB => 3 x 12 GB modules

high-end M4 Max: 128 bits/module x 4 modules x (8533 x 10^6)/s x 1 byte/(8 bits) x 1 GB/(10^9 bytes) = 546.112 GB/s
Memory options: 48 GB => 4 x 12 GB modules; 64 GB => 4 x 16 GB modules; 128 GB => 4 x 32 GB modules

 

[NotEntirelyConfused]

It looks like they went back to M1/2 family system. M3 so far is an outlier with a weaker Pro. I guess the customers didn’t like it that much?

M4 Pro is a bit stronger vs. M4 than the M1 or M2 Pros were vs. their base chips.

 

[theorist9]

I think for most people buying the Mini, more cpu is probably more useful than more gpu. I would love huge gpu improvements personally, but that doesn’t seem to be the case for the majority.

I think that's true for now. But if a sufficent AAA gaming ecosystem ever develops for MacOS, a GPU-heavy Mini could become a popular machine.

 

[Jimmyjames]

One curious thing about the Cyberpunk 2077 announcement. They mention frame generation in the press release. To my knowledge, that isn’t currently part of MetalFX. I believe FSR 3.1 has it? So is MetalFX getting an upgrade to allow this? Will they just be using FSR?

 

[casperes1996]

For easier comparison, I've updated my Mx table. Now I wonder if we are ever going to see an M3 Ultra...
EDIT: Fixed a typo.
EDIT2: Updated table with M4 Max is on the next page. Outdated table removed.

M3 Ultra definitely never happening. M4 Ultra, - yes.

As for the core count on Pro - Mac are amazingly popular among software engineers, where the Pro is the most common chip (at least at my company) - The CPU cores deliver, the GPU cores sleep. For personal hobby stuff at home, I want more GPU less CPU, for work (and WFH) I want as much CPU as you can give me