iOS 18 Leak Reveals Apple Tested MacBook Pros With M3 Ultra Chip

[tenthousandthings]

~

This could be very well the case. Where the M1 and M2 Ultras are just 2 maxes together, the M3 Ultra has different RAM with more density. I’m not sure if there are other differences, but that means if the M3 Ultra still is 2x M3Max, it’s actually M3Max rev2 x2.

Exactly. Another difference is Thunderbolt 5 support, which is on the chip(s). Not to mention the UltraFusion interconnect(s).

While it’s interesting to speculate about what an Ultra MacBook Pro would look like, this leaker has misinterpreted the “d” suffix, in two different ways:

1. It may not only indicate the M3 Ultra, it may also indicate the M3 Max rev2, as follows:

J514 = M3 Pro/Max MacBook Pro 14"
J516 = M3 Pro/Max MacBook Pro 16"
J514d = M3 Max rev2 MacBook Pro 14"
J516d = M3 Max rev2 MacBook Pro 16"
J614 = M4 Pro/Max MacBook Pro 14"
J616 = M4 Pro/Max MacBook Pro 16"

J575 = M4 Max Mac Studio
J575d = M3 Ultra Mac Studio

2. The J “codenames” indicate the chassis. So, for example, J514 is the third-generation Pro/Max MacBook Pro 14" chassis. J575 is the third-generation Max/Ultra Mac Studio chassis.

The “d” suffix indicates the presence of this revision of the M3 Max, either by itself or in the M3 Ultra. The leaker misses the (probable) fact that, if Apple had been testing an M3 Ultra MacBook Pro to go alongside the M4 Pro/Max MacBook Pro in the fourth-generation chassis, it would be J614d and J616d, which is not what the leaker found.

 

[tenthousandthings]

If the MBP can handle the Ultra (or even if they were only considering it), then that confirms we got screwed on the Mac Studio M3 Ultra for sure. Should've been M4 Ultra.

If an M4 Ultra Mac Studio existed, it would be J575, not J575d. But it doesn’t exist, and it will never exist, because M5 Ultra has displaced it. By skipping M4 Ultra, Apple can advance M5 Ultra on TSMC N3P earlier in the development process. This pattern is likely to continue going forward, with M7 Ultra on TSMC N2P (skipping M6 on TSMC N2), and so on.

The M3 Ultra Mac Studio is a stop-gap solution, for a year at most. Far from getting “screwed,” these users have been given an interim option that Apple has gone out of its way to provide.

 

[Frantisekj]

The New PowerBook G5!

It looks some employee wanted to make waffle 🧇 at work 😂

 

[darngooddesign]

I think your reading comprehension is lacking.

Not as lacking a the understanding of what’s required to cool an M#-Ultra under load, let alone one in a MacBook Pro. I mean, we literally know that in the studio it’s heatsink weighs 2 pounds more than the Max’s, and that’s with the huge fan required the keep it quiet and cool.

 

[redheeler]

The New PowerBook G5!

View attachment 2513243

People might finally stop complaining about Apple's 'obsession with thinness'.

Only on MacRumors are the memes old enough to legally buy a drink!

 

[profcutter]

Well said. Much like those that cried about the discontinuation of the 17" MacBook. They made a big stink but in truth, sales of the 17" model were less than 1% of Apple laptops sold. No reason to keep around a product that sells so few units.

What people say and what they do is two different things. Those who complain about the Mac Pro upgrades aren't going to buy a Mac Pro no matter what (unless they cut the price by 90%). They aren't the target market but they act like they are and are incredibly vocal about it.

I mean, people say that, but I definitely had a 17”, and folks I knew in creative fields had them too.

 

[deconstruct60]

This could be very well the case. Where the M1 and M2 Ultras are just 2 maxes together, the M3 Ultra has different RAM with more density. I’m not sure if there are other differences, but that means if the M3 Ultra still is 2x M3Max, it’s actually M3Max rev2 x2.

The M3Max rev2 wafers would come with defects. Some dies will have defects in the UltraFusion subsection. So they won't be useful in a x2 package. It would be cheaper , faster, and simpler to drop then into a single die package ( if the die hasn't grown too big. ).

Seems like more a need for a 'test mule' chassis than a product development line. Throw 'half node' M3Max rev2 at some developers to work on the software stack while wait for the volume of. dual die packages to build up to something substantive. Way back at the the start of N3B planning "Max rev2" might have been "Plan A", but that didn't come to fruition.

 

[bigglow]

Apple should do this, even if they would be immensely expensive and suffer for limited battery. This would still be amazing halo product that would destroy any comparisons to competition.

Add a ultra-only limited subtle color in there and this will become a magical status symbol in any cafe

There's a market for it anyways a demonstrated by how many Intel HX laptops

 

[rp2011]

M3 Ultra chips are not very good. Or at least a good value. The Max is a far better solution all around. Even for the Studio, the Ultra chips have been very disappointing, M3 Ultra included.

 

[jole]

M3 Ultra chips are not very good. Or at least a good value. The Max is a far better solution all around. Even for the Studio, the Ultra chips have been very disappointing, M3 Ultra included.

That statement makes little sense:

M3 Ultra gives you ~2x multithreaded performance, 4x max memory, ~2x GPU. If you need those, the higher price tag might be worth it.

If you do not need what M3 Ultra provides over M4 Max, go for Max.

 

[rp2011]

That statement makes little sense:

M3 Ultra gives you ~2x multithreaded performance, 4x max memory, ~2x GPU. If you need those, the higher price tag might be worth it.

If you do not need what M3 Ultra provides over M4 Max, go for Max.

The performance of the M4 is so good that it makes the performance difference between the 3 Ultra and 4 Max not too impressive. There is no reason for the M3 Ultra to exist; everyone should skip 3 Ultra and wait for the next iteration. Not worth it. Not good.

 

[izzy0242mr]

Just waiting for the proper spin on 2 hrs SoT as "all-day battery life"

A "day" is relative. There's probably some planetoid out there that spins very quickly and has a 2-hour "day" cycle. When Apple finds it they'll be able to make this claim!

 

[izzy0242mr]

Apple could build a MacBook that uses 240 Watts and has crazy specs and a 16 inch screen. It would power down 3/4 of the CPU and GPU cores when on battery and run the full system only when plugged into power. Lots of people say they want this, but VERY FEW would actually pay for it.

This would be a pretty marked departure from the norm for Apple, no? I don't even think that their Intel Macs switched to lower-performance modes when not on AC power. Obviously all chips can throttle if overheating, but while Windows PCs typically do change by default how fast your computer functions when you go from AC power to battery, macOS and most versions of Mac OS X to my knowledge have never done this.

So having a laptop that, for the first time, gives poorer performance when on battery would be a negative move for Apple from a PR perspective because it results in a laptop that has inconsistent performance. They didn't even have a "Low Power Mode" until recently with Apple Silicon, and even that is (a) optional and (b) is based on battery being low, not merely using the battery vs being plugged in.

Apple—in its current form with its current ethos—will never make a laptop that has different performance modes when plugged in vs on battery.

 

[jole]

The performance of the M4 is so good that it makes the performance difference between the 3 Ultra and 4 Max not too impressive. There is no reason for the M3 Ultra to exist; everyone should skip 3 Ultra and wait for the next iteration. Not worth it. Not good.

4x RAM, 2x compute is not impressive?

With 512GB RAM M3 Ultra is the only game in town for running local inference for 600B+ param models. Similarly, it allows you train drastically larger models than anything else on the market.

And for compute, ~ 2x is not exactly a small difference.

 

[rp2011]

4x RAM, 2x compute is not impressive?
With 512GB RAM M3 Ultra is the only game in town for running local inference for 600B+ param models. Similarly, it allows you train drastically larger models than anything else on the market.

And for compute, ~ 2x is not exactly a small difference.

It's not 2 times compute. 1. The Ultras do not scale. 2. The M4 Max single-core scores blow it away. These are chips you should skip for the next version. The performance difference between the M4 Max and the 3 Ultra is marginal. The Ultras have always been disappointing. If someone is desperate enough to use a 3 Ultra to run a smaller AI model locally, because of the memory available to them, that's one use case.

 

[jole]

It's not 2 times compute. 1. The Ultras do not scale. 2. The M4 Max single-core scores blow it away. These are chips you should skip for the next version. The performance difference between the M4 Max and the 3 Ultra is marginal. The Ultras have always been disappointing. If someone is desperate enough to use a 3 Ultra to run a smaller AI model locally, because of the memory available to them, that's one use case.

If you plan using single-core, do not pay $10k for a 32-core Ultra

 

[rp2011]

If you plan using single-core, do not pay $10k for a 32-core Ultra

That was an example of single core, in multi core, in most tests the difference between the M4 Max and M3 Ultra were marginal. But believe what you want. It’s your money. There are plenty of tests out there comparing the two. Check for yourself.

 

[jole]

It's not 2 times compute. 1. The Ultras do not scale. 2. The M4 Max single-core scores blow it away. These are chips you should skip for the next version. The performance difference between the M4 Max and the 3 Ultra is marginal. The Ultras have always been disappointing. If someone is desperate enough to use a 3 Ultra to run a smaller AI model locally, because of the memory available to them, that's one use case.

Browsed some benchmarks — it is complicated:

• M4 cores are significantly faster than M3 cores. Theoretically maxed out 32 core M3 Ultra should be 50% faster than maxed out 16 core M4 Max on multithreaded CPU tests. Many benchmarks show way less than that.
• On GPU M3 Ultra with 80 cores should eat 40 core M4 Max for lunch — the cores in M3 Ultra are updated from M3. Many benchmarks do not reflect this 2x expectation for some reason. Maybe this is memory bandwidth limited as the Ultra M3 is has 50% more theoretical bandwidth, not 100%
• Memory is the top advantage of M3 Ultra. If you need more than 128GB, Ultra is the only game in town. Also, 50% more bandwidth is a significant jump.
• Single core performance can often be the most important thing. Max dominates there.

All-in-all, Ultra is very specialized machine for special use-cases.

Curious to see if Apple will be able to fix the bottlenecks for AI workloads. While M3/M4 show promise, they are still quite limited when it comes to matrix multiplications (especially inference with quantized models) compared to NVIDIA. They seem to be working on it for AI servers at least.

 

[rp2011]

Browsed some benchmarks — it is complicated:

• M4 cores are significantly faster than M3 cores. Theoretically maxed out 32 core M3 Ultra should be 50% faster than maxed out 16 core M4 Max on multithreaded CPU tests. Many benchmarks show way less than that.
• On GPU M3 Ultra with 80 cores should eat 40 core M4 Max for lunch — the cores in M3 Ultra are updated from M3. Many benchmarks do not reflect this 2x expectation for some reason. Maybe this is memory bandwidth limited as the Ultra M3 is has 50% more theoretical bandwidth, not 100%
• Memory is the top advantage of M3 Ultra. If you need more than 128GB, Ultra is the only game in town. Also, 50% more bandwidth is a significant jump.
• Single core performance can often be the most important thing. Max dominates there.

All-in-all, Ultra is very specialized machine for special use-cases.

Curious to see if Apple will be able to fix the bottlenecks for AI workloads. While M3/M4 show promise, they are still quite limited when it comes to matrix multiplications (especially inference with quantized models) compared to NVIDIA. They seem to be working on it for AI servers at least.

Exactly. I told you

 

[tenthousandthings]

Browsed some benchmarks — it is complicated:

• M4 cores are significantly faster than M3 cores. Theoretically maxed out 32 core M3 Ultra should be 50% faster than maxed out 16 core M4 Max on multithreaded CPU tests. Many benchmarks show way less than that.
• On GPU M3 Ultra with 80 cores should eat 40 core M4 Max for lunch — the cores in M3 Ultra are updated from M3. Many benchmarks do not reflect this 2x expectation for some reason. Maybe this is memory bandwidth limited as the Ultra M3 is has 50% more theoretical bandwidth, not 100%
• Memory is the top advantage of M3 Ultra. If you need more than 128GB, Ultra is the only game in town. Also, 50% more bandwidth is a significant jump.
• Single core performance can often be the most important thing. Max dominates there.

All-in-all, Ultra is very specialized machine for special use-cases.

Curious to see if Apple will be able to fix the bottlenecks for AI workloads. While M3/M4 show promise, they are still quite limited when it comes to matrix multiplications (especially inference with quantized models) compared to NVIDIA. They seem to be working on it for AI servers at least.

Hence skipping M4 Ultra and going to M5. Should be interesting, maybe we’ll know more June 9.