M4 Mac Announcements

[NotEntirelyConfused]

SFP+ moves the transceiver to a daughterboard that plugs into the port. So you can either use a transceiver that exposes RJ-45, or an optical connection for standard optical cables, or even fully integrated cables. I bet data centers hosting Macs would love it.

Yes exactly. In DCs, using DACs instead of 10BaseT is a big win for power (and cooling the switch). Right up Apple's alley, really. And for non-DC uses, having the flexibility of using SFPs is great. Right now you'd need a shitty media converter to use fiber, or you'd have to ignore the built-in Ethernet entirely and hang something off a TB port.

 

[NotEntirelyConfused]

4 things were listed (all as the "full fat" variants), you need to be more specific on which "it" you are referring to here. I have to assume you mean M2 Pro here?

Yes, apologies, I have already edited my post.

Considering the M4 Pro already has two variants: 8+4 and 10+4, I'd be very surprised if it was a cluster of 6. That would mean fusing off 4 cores for the 8+4 Mac mini being offered for pre-order right now.

Yes, I agree, that's why I think you're right about 15P. (Again, assuming door #3, which isn't a given.)

 

[theorist9]

SFP+ moves the transceiver to a daughterboard that plugs into the port. So you can either use a transceiver that exposes RJ-45, or an optical connection for standard optical cables, or even fully integrated cables. I bet data centers hosting Macs would love it.

As I said, I'm not familiar with this stuff, so I'm not following this explanation. I looked up "transciever" for fiber optics, and it explained that's the term for the opto-electronic converter. But I still don't know what "transceiver that exposes RJ-45" means, etc., etc.

 

[casperes1996]

You're both wrong. It was 6/8P and 4E.

Yes, sorry. Correct list

M1 Pro: 8+2
M2 Pro: 8+4
M3 Pro: 6+6
M4 Pro: 10+4

Even less pattern!

 

[ArgoDuck]

Yes, sorry. Correct list

M1 Pro: 8+2
M2 Pro: 8+4
M3 Pro: 6+6
M4 Pro: 10+4

Even less pattern!

"I am ill at these numbers." (Hamlet).

 

[NotEntirelyConfused]

As I said, I'm not familiar with this stuff, so I'm not following this explanation. I looked up "transciever" for fiber optics, and it explained that's the term for the opto-electronic converter. But I still don't know what "transceiver that exposes RJ-45" means, etc., etc.

Try this wikipedia article.

Basically, it's a good way to abstract out the media-specific hardware from everything else. The Ethernet chip does all the work, except the MAC part, which goes in the SFP.

Past 10Gbps, it gets slightly more complex, as you have a bunch of different transceiver types, some of which allow multiple Ethernet ports per socket. The very latest (OSFP) allows a single 800gbps, dual 400gbps ports, or 4x200Gbps.

 

[Nycturne]

As I said, I'm not familiar with this stuff, so I'm not following this explanation. I looked up "transciever" for fiber optics, and it explained that's the term for the opto-electronic converter. But I still don't know what "transceiver that exposes RJ-45" means, etc., etc.

Transceiver is just a component that transmits and receives over a medium via some form of encoding for that medium, it is short for "Transmitter/Receiver". The medium doesn't matter really: radio/wifi, optical, coax, twisted pair, USB-C, etc...

Encodings have gotten more complex as we try to pack more information, compared to say, the simple encoding of a serial port that uses TX/RX wires.

But the idea is that instead of baking the transceiver into the Ethernet PHY, you can put it on a small board. This board handles translating to the medium (twisted pair or single/multi-mode optical cables for example) that will carry the raw signal.

 

[Andropov]

M1 Pro: 8+2
M2 Pro: 8+4
M3 Pro: 6+6
M4 Pro: 10+4

The new Pro has got to be a significant upgrade over the M3 Pro. The M2 Pro -> M3 Pro upgrade was modest at best due to the reduced number of P cores, this reversal of the change will have the opposite effect.

Interesting pattern in any case. I wonder what kind of data Apple engineers look at while choosing a number of E cores for the CPU.

 

[casperes1996]

Interesting pattern in any case. I wonder what kind of data Apple engineers look at while choosing a number of E cores for the CPU.

I think they just roll a D10

 

[Andropov]

SFP+ moves the transceiver to a daughterboard that plugs into the port. So you can either use a transceiver that exposes RJ-45, or an optical connection for standard optical cables, or even fully integrated cables. I bet data centers hosting Macs would love it.

This reminds me that I wish Apple hadn't discontinued the Xserve.

 

[Jimmyjames]

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.

I’d love 3. It has to be 15 or 18 P cores surely?

I wonder if it would be too much (thermally) for the MBP chassis? Can’t deny 18 P cores would be mind blowing though.

 

[NotEntirelyConfused]

I’d love 3. It has to be 15 or 18 P cores surely?

I wonder if it would be too much (thermally) for the MBP chassis? Can’t deny 18 P cores would be mind blowing though.

This topic has also come up at MR. I'm not so sure.

The same type of argument could be made against the massive many-core server chips from AMD, Intel, etc. And yet... because of the shape of the perf/power curve, they find significant value in running all their cores at roughly half the speed they could, if they were just running a few of the cores.

Apple's efficiency is vastly better than theirs, but the shape of the curve isn't *that* different. So a 10P M4 Pro running at N watts is likely to be significantly less performant that a 20P M4 Max running at the same N watts. That's assuming a workload that is parallel enough, which is not going to be the case, lots of the time. But when it is, the many-cores Max wins.

Honestly, I don't think Apple is going to go that high. (And in fact, this might well explain the rumor that they're doing a totaly new design for Ultra - a desktop-only chip might well justify that many-core argument, where a laptop chip doesn't.) It doesn't seem impossible though.

 

[mr_roboto]

Oh absolutely, but having them be the same across these three family members appeals to my sense of consistency. That’s not necessarily a good reason, but a better one might be that sharing the same cluster design across multiple dies might decrease the overall workload for the Apple engineers.

Maybe there are actually two clusters with 6 cores each, but one is disabled. We will see how M4 Max looks like and whether it is a chopped die or a fully separate design. If M4 Max has 18 performance cores, that will literally embarrass the rest of the industry.

Considering the M4 Pro already has two variants: 8+4 and 10+4, I'd be very surprised if it was a cluster of 6. That would mean fusing off 4 cores for the 8+4 Mac mini being offered for pre-order right now.

I think you're all barking up wrong trees.

You're not considering another far more likely possibility. A CPU cluster is a hierarchy of blocks, not a single flattened block. Apple should have some flexibility in designing variants with fewer or more cores, and this shouldn't require them to redo all physical design work for each variant.

To illustrate, I'm attaching High Yield's annotated die photos of the whole M3 family. They show that Apple designed a 6-core P cluster, but deleted two cores from it for the base M3, using the freed-up area for other circuits. They also show layout differences between the 4- and 6-core variants of the E cluster which go slightly beyond just deleting two cores from the 4-core variant (two of the cores shift location slightly).

 

[Altaic]

Huh? I can't think of a single ProMotion display that's 240Hz max, Apple's only been doing 120Hz.

Huh, for some reason I thought it topped out at 240Hz. Thanks for the correction!

 

[Altaic]

By the way, there is something very odd about the M4 Pro RAM bandwidth. The M4 is quoted at 120GB/s, that’s the usual LPDDR5X. But M4 Pro is whopping 273GB/s, more than double! I have difficulty understanding which memory technology is that. If the RAM standard is the same, this would indicate a 320-bit interface with ECC. If it’s still a 192- or 256-bit interface, then it must be some new RAM tech. This is too fast even for LPDDR6.

I predicted 240 or 273 GB/s earlier in this thread (see below). Ended up getting everything I was hoping for, too. The M4 lineup is going to be absolutely stellar.

Something I haven't seen anyone else mention is that the M4 Pro has a max of 64GB RAM, which would mean an even number of memory channels. The M3 Pro had 3x 64bit channels, which disappointed some people, as memory bandwidth was 3/4 of what was expected with 4x 64bit or 2x 128bit that were in M1/2 Pro generations.

The M4 Pro therefore should have 1.33x bandwidth from that, plus the uplift from LPDDR5-6400 going to LPDDR5X-7500 (1.17x) or LPDDR5X-8533 (1.33x), so 240 GB/s or 273 GB/s. I was really hoping for this given my work in ML, so I'm very excited. Also really hoping for TB5... Fingers crossed.

 

[leman]

I predicted 240 or 273 GB/s earlier in this thread (see below). Ended up getting everything I was hoping for, too. The M4 lineup is going to be absolutely stellar.

Absolutely! The M4 pro is essentially as fast as my M3 Max (or even faster!) Only the GPU is still a bit lackluster. I hope Apple has some nice GPU upgrades planned for the next years.

 

[Cmaier]

I think the max is going to be surprising, at least compared to what we thought it was going to be as of yesterday.

 

[dada_dave]

Absolutely! The M4 pro is essentially as fast as my M3 Max (or even faster!) Only the GPU is still a bit lackluster. I hope Apple has some nice GPU upgrades planned for the next years.

I haven’t admit, if forced to choose, I would’ve preferred them to throttle back the CPU core count and increased the GPU count for the Pro die. But let’s see what the Max has in store. I have the feeling Apple is still experimenting with different configurations. Unfortunately the next big lithography upgrade won’t be until 2026/M6. N3P will be nice I’m sure for M5 but not a major density increase. That said depending on how well TSMC executes we might get two density increases in a row with N2 (2026/M6) and A16 (2027/M7). So I don’t know if GPU core counts will change that much for the M5. Of course that says nothing about architectural changes and Apple could increase core counts anyway.

 

[theorist9]

By combining two external display controllers, the M4 & M4 Pro can drive 8k@60 Hz. If the M4 Max has, like the M3 Max, one internal controller and four external controllers*, they could, in principle, combine the latter to drive 8k@120 Hz. And since the M4's TB4 can support 8k@60, the M4 Max's TB5 could support 8k@120.

*Or it might just have five externals since, based on display support, the M4 Pro appears to have 3 externals rather than 2 externals + 1 internal (i.e., doing away with the internal controller entirely). [The distinction between external and internal is purely based on capacity. The lower-capacity (and supposedly more efficient) internal controllers, which are designed for the built-in screens, can also be used to drive external displays. E.g., while the M4 Mini appears (again, based on display support) to have 2 externals + 1 internal, all three are used to drive external displays.]

 

[theorist9]

BTW, here's the US pricing comparison for the M4 vs. M2 Mini.

Comparing pricing on devices with minimum RAM (which is higher for the M4 than the M2), and the the same storage:

M4 Mini vs M2 Mini: Same price.

M4 Pro Mini vs. M2 Pro Mini, lower-end Pro chip: M4 is $100 more.

M4 Pro Mini vs. M2 Pro Mini, higher-end Pro chip: Same price.