M1 Ultra is the last chip based on the M1

[mr_roboto]

There is nothing that says Apple couldn’t have a quad M1 Max.

The people who have been reverse engineering M1, M1 Pro, and M1 Max to port Linux to them found, several months ago, that (a) the interrupt controller in M1 Pro/Max supports multi-die configs and (b) the possible configurations are one die and two die. No support for four die.

There is no reason to doubt that information based on what Apple disclosed today.

 

[crazy dave]

I would say the "final member of the M1 family" doesn't mean final member of the "M1 architecture."

What would not be a member of the M1 family but still a member of the M1 architecture? That doesn't make any sense. You're thinking they'll rebrand the workstation chips? Which base dies will they use then? Because it doesn't seem like it will be the M1 Max die ... not as we've seen it anyway.

Because that's how the industry works. Apple is doing it, AMD and Intel have been doing it for decades.

It takes time to develop and build around the A14 Firestorm and Icestorm cores. It takes a couple years to enhance the A15 Avalanche and Blizzard cores.

Workstation chips always come last, desktop and mobile first. It's only today, March 8 that AMD is launching Zen 3 based workstation chips. The architecture was launched in November 2020, same time as M1.

Workstation and desktop chips have indeed come out before low power, smaller chips. Graphics cards do that regularly - in fact, sometimes the smallest, lowest power chips are the very last to come out. For Alder Lake, Intel went desktop first. Supposedly AMD will too for Zen 4. Threadripper may have taken forever to come out but that's because AMD has been focusing on even larger server chips.

That said, I think Apple will introduce the base M2 first and there is another option here: one of the weirder rumors is that the Mac Pro will remain Intel for now, even getting an upgraded Intel chip this year. That makes the whole "transition will be completed in two years" statement a little dubious to the say the least so I am far from certain about this, but it's possible that it may be awhile before we see a full Apple Silicon Mac Pro.

 

[Andropov]

That said, I think Apple will introduce the base M2 first and there is another option here: one of the weirder rumors is that the Mac Pro will remain Intel for now, even getting an upgraded Intel chip this year. That makes the whole "transition will be completed in two years" statement a little dubious to the say the least so I am far from certain about this, but it's possible that it may be awhile before we see a full Apple Silicon Mac Pro.

They also said that they had more Intel Macs to be released. At this point, I'd say one of those two statements may have been wrong.

 

[crazy dave]

They also said that they had more Intel Macs to be released. At this point, I'd say one of those two statements may have been wrong.

Some people have pointed out that technically an Intel iMac was released post-sending out of dev kits but before M1 Macs were actually released. My impression was that transition started with the release of the Macs so I also feel that intel iMac release is still ... unsatisfying as to the timing. But yes something has to give here.

 

[JPack]

What would not be a member of the M1 family but still a member of the M1 architecture? That doesn't make any sense.

A hypothetical "W1" chip for the workstation with user-accessible DIMM slots, based on the Firestorm and Icestorm cores. What Apple calls "M1 family" is all marketing. If Apple refreshes the Mac Pro with such a chip, I doubt they're going to call it "M1 Ultra Extreme." It'll have a proper workstation marketing name.

Intel Core/Xeon
AMD Ryzen/Threadripper Pro
Apple M1/W1

Workstation and desktop chips have indeed come out before low power, smaller chips. Graphics cards do that regularly - in fact, sometimes the smallest, lowest power chips are the very last to come out. For Alder Lake, Intel went desktop first. Supposedly AMD will too for Zen 4. Threadripper may have taken forever to come out but that's because AMD has been focusing on even larger server chips.

That said, I think Apple will introduce the base M2 first and there is another option here: one of the weirder rumors is that the Mac Pro will remain Intel for now, even getting an upgraded Intel chip this year. That makes the whole "transition will be completed in two years" statement a little dubious to the say the least so I am far from certain about this, but it's possible that it may be awhile before we see a full Apple Silicon Mac Pro.

Intel and AMD both focus on launching chips based on strategic need and margins. Intel is very well established with Xeon. AMD not so much so we see them with EPYC sooner. Apple is virtually non-existant in the workstation and datacenter space and they aren't about to take over, so why would Apple develop M2-based workstation chips first? Their focus is on consumer and pro-sumer devices.

 

[Juuro]

What we can confirm is that UltraFusion I a multi-die interconnection archcture so its not impossible that it will be something like this ... because UltraFusion and multi-die seems strange only to waste on one product...

Mac Studio
M1 Ultra Chip (2 M1 Max Chip) using UltraFusion
20 CPU Core
64 GPU Core
128 GB Unified memory

Mac Pro
2xM1 Ultra Chip (4 M1 Max Chip) using UltraFusion
40 CPU Core
128 GPU Core
256 GB Unified memory

Mac Pro
4xM1 Ultra Chip (8 M1 Max Chip) using UltraFusion
80 CPU Core
256 GPU Core
512 GB Unified memory

Mac Pro
6xM1 Ultra Chip (12 M1 Max Chip) using UltraFusion
120 CPU Core
384 GPU Core
768 GB Unified memory

If Apple didn't lie in their keynote today that is simply not possible.

The M1 Max has some kind of interface on one (!) side. There was a developer who found that already in October 2021. Today Apple confirmed his findings with their announcement that they use two M1 Max dies to form the M1 Ultra. UltraFusion works with two dies and only two dies because M1 Max has only one interface for something like that.

Also John Ternus said that with the M1 Ultra they add one last chip to the M1 family. If two M1 Max dies connected with UltraFusion qualifies as a chip four M1 Max dies would qualify as another chip and such a thing would be another chip in the M1 family. So that can not happen because John Ternus said M1 Ultra is the last chip of the M1 family.

For the Mac Pro I can imagine two things:
1. They use multiple M1 Ultras SoCs together in Mac Pro connected via the motherboard. I'm sceptical of this solution because the kind of discredited it in todays event. But they could argue that two M1 Ultras would still be much more powerful than only one despite the tradeoffs. Also the worse power consumption is not much of an issue with the Mac Pro.
2. They wait for M2 Ultra and there we will have a possibility for 4 dies together. Together with the rumour of Kuo that Mac Pro will come next year I think at the moment that this is the most likely possibility.
3. They come up with a new better unheard of architecture of connecting two SoCs.

 

[crazy dave]

A hypothetical "W1" chip for the workstation with user-accessible DIMM slots, based on the Firestorm and Icestorm cores. What Apple calls "M1 family" is all marketing. If Apple refreshes the Mac Pro with such a chip, I doubt they're going to call it "M1 Ultra Extreme."




Intel and AMD both focus on launching chips based on strategic need and margins. Intel is very well established with Xeon. AMD not so much so we see them with EPYC sooner. Apple is virtually non-existant in the workstation and datacenter space and they aren't about to take over, so why would Apple develop M2-based workstation chips first? Their focus is on consumer and pro-sumer devices.

Except Avalanche and Blizzard cores will be over a year old by then and the node to produce them fully mature. There's no reason they couldn't have designed their product roadmap with that in mind. It's also not clear to me that Apple will have the same Xeon/Epyc divide with Apple silicon. Maybe you're right, but it certainly isn't going to be 4xM1 Max dies then, it would have to be whole new dies, which doesn't seem likely given what we know so far.

(EDIT) There is one possibility: there is a rumored 12-core CPU die that we haven't seen yet. Maybe that's still firestorm/icestorm architecture. If that's accurate and they rebrand it a "W" chip or something and that can be made into 4x dies then maybe.

 

[mr_roboto]

We'll see when somebody decaps it and sees if its a single piece of silicon or not. I rewatched the relevant portion and he said "it connects over 10000 signals", which is vaguely ambiguous. If its 10K wires, it really suggests the interconnect is on the wafer itself between two M1 MAX die. If its referring to an array of SERDES channels that bundles "signals" as packets, serializes them, and deserializes them on the other side, then it seems there's IP out there for that.

I'm not sure how you've convinced yourself there's a chance it's a single die when Apple clearly stated that it is not, showed off the interconnect rectangle at the bottom of the M1 Max die which they didn't disclose at M1 Max launch, said the interconnect technology is silicon interposer (which is compatible with ~10K wires), and M1 Max is already such a large die that it would be ridiculous to try to yield something twice as big.

Maybe you're not familiar with silicon interposers? AFAIK they were first commercialized in Xilinx 7 series FPGAs about 10 years ago. The idea is that you build an ultra high density wiring substrate - an interposer - out of passive silicon in an older process node (iirc, for Xilinx 7 series it was 65nm interposer, 28nm logic). The interposer uses TSVs so it can make solder bump connections to both the organic package substrate on the bottom of the stack and logic die on the top.

The silicon interposer can be quite huge relative to active devices since there's no active logic. Much higher yields and dramatically fewer process steps than an active device. Assembly yields can bite you, IIRC they did early on with those big Xilinx 7 series FPGAs, but presumably that's a solved problem by now.

 

[crazy dave]

They also said that they had more Intel Macs to be released. At this point, I'd say one of those two statements may have been wrong.

You know the Mac Pro did get a new GPU option today ... maybe that's what they'll count as the new intel-based product? ? I dunno, something has to give.

 

[killawat]

2. They wait for M2 Ultra and there we will have a possibility for 4 dies together. Together with the rumour of Kuo that Mac Pro will come next year I think at the moment that this is the most likely possibility.

I agree with this. 2x 20-Core CPU, particularly how Apple has optimized everything, would have worse performance characteristics than a 40-Core CPU. Same with GPU and Neural Engine. Apple has a great thing going with the existing pipeline. Given the fantastic performance of the M1 Ultra, I don't think we'll need to wait for the M2 in the Mac Pro. I will go out on a limb and say that most Mac Pro users would be satisfied with the M1 Ultra as a drop in supplement to the existing line up (save for compatibility and intel dependence). The Mac Pro could continue to have PCIe or whatever the heck they're calling it these days but may not offer Radeon GPU compatibility. Considering Apple hasn't changed their stance on lower end eGPUs and even second part solutions like the BlackMagic eGPU.

 

[JPack]

Except Avalanche and Blizzard cores will be over a year old by then and the node to produce them fully mature. There's no reason they couldn't have designed their product roadmap with that in mind. It's also not clear to me that Apple will have the same Xeon/Epyc divide with Apple silicon. Maybe you're right, but it certainly isn't going to be 4xM1 Max dies then, it would have to be whole new dies, which doesn't seem likely given what we know so far.

Apple has limited silicon engineering resources and has to decide on priorities. I'm sure in Tim Cook's dream roadmap, the entire M1 architecture worth of products would all come out in one big bang on November 10, 2020. Right now, it has taken 483 days to launch M1 through M1 Ultra. If we reverse that, when would M2-based MacBook Air or iMac come out?

Why would it be entirely new dies? They can use current MCM to duplicate M1 Ultra. It's not ideal, but it works. Based on what we know, something call Jade 4C-Die is coming. That doesn't sound like a whole new chip to me.

 

[crazy dave]

Apple has limited silicon engineering resources and has to decide on priorities. I'm sure in Tim Cook's dream roadmap, the entire M1 architecture worth of products would all come out in one big bang on November 10, 2020. Right now, it has taken 483 days to launch M1 through M1 Ultra. If we reverse that, when would M2-based MacBook Air or iMac come out?

Why would it be entirely new dies? They can use current MCM to duplicate M1 Ultra. It's not ideal, but it works. Based on what we know, something call Jade 4C-Die is coming. That doesn't sound like a whole new chip to me.

To be clear, I don't think the M2 Air will be last out of the door. It's not clear to me yet what the Mac Pro is going to be or when.

The problem is to replace the Mac Pro with what you are describing (user replaceable DIMMs, multiple PCIe slots, etc ...), the AS dies have to have significantly more IO than the Max, even at 4x, can provide. It also probably need a new memory controller/interface and a new IRQ. The current Ultra *is* an MCM already and the M1 Max can only support up to two dies connected together with its IRQ. If Apple is to release a "W1" based on firestorm/icestrom with all of that ... that has to be a new base die. Maybe the rumored 12 core die? Maybe not.

 

[Fomalhaut]

Because 2x32 is significantly more expensive than 2x24?

Yes, but the M1 Max 24 to 32 core upgrade is $200, so 2 x M1 Max upgrades should be $400, not $1000.

If we assume the M1 Ultra has two M1 Max dies that are then "joined", it looks the majority of the cost is actually the "silicon interposer" that does the joining, rather than the GPU core upgrades.

However, you are already paying this cost with the jump from the M1 Max to even the binned (48-core GPU) Ultra, so I don't know where the $600 "mark-up" for the GPU cores comes from.

Considering the binned Ultra model has 1TB SSD and 64GB RAM, the upgrade cost to the entry-level Studio Ultra is $1400, which considering you are doubling up on CPU & GPU cores, neural engine, memory bandwidth is actually not too bad.

I expect that Apple will sell a lot more of the base $3,999 Studio Ultras (with some SSD upgrades) than the maxed out 64-core GPU, 128GB RAM versions that add $1800 to the price.

 

[mr_roboto]

Why would it be entirely new dies? They can use current MCM to duplicate M1 Ultra. It's not ideal, but it works. Based on what we know, something call Jade 4C-Die is coming. That doesn't sound like a whole new chip to me.

Based on what we know? Jade 4C-Die was an early rumored codename, but now there are real physical objects to examine and rumors don't hold as much weight any more. It's been mentioned several times already: the people who are reverse engineering Apple Silicon to port Linux found that the M1 Pro/Max interrupt controller had support for 2 die. Not three die, not four, two.

You were right to say that Apple has to decide on priorities. I'm not sure why you aren't considering the possibility that one such priority call was simply not addressing the Mac Pro product segment in the M1 family of chips.

Consider that Apple announced the transition would take 2 years. It's clear they've taken a bottom-up approach, so that puts the Mac Pro last. If their planners knew it would ship 2 full calendar years after the first M1 product launch, why not try to at least make the first-gen Mac Pro silicon a member of the M2/A15 technology generation, if not the M3/A16? It would look slightly silly if the brand new Mac Pro chip was 2 generations behind in single thread CPU performance.

In the above, I am assuming they plan to make a M series derivative of A15. But even that's not a given. Another priority call they may have made is to put Macs on a 2 year silicon cadence instead of 1 year like phones. They already were doing a pattern like that with high end tablet chips - iPad Pro went from A10X to A12X/Z to M1 (which would've been called A14X if not for putting Apple Silicon in the Mac). There was no A11X or A13X.

 

[bobcomer]

Alder Lake is not a workstation product.

From what I see it is, but then you probably have a different definition for workstation. Alder lake has all the normal intel types, ultramobile, mobile, NUC, SFF Desktop, desktop, workstation.

 

[bobcomer]

Count the cores on Alder Lake then count the cores on M1 Ultra.

Okay, 14 cores (with HT) for Alder Lake, 20 for the Ultra. And that doesn't make any difference whether it's a workstation or not.

 

[JPack]

Based on what we know? Jade 4C-Die was an early rumored codename, but now there are real physical objects to examine and rumors don't hold as much weight any more. It's been mentioned several times already: the people who are reverse engineering Apple Silicon to port Linux found that the M1 Pro/Max interrupt controller had support for 2 die. Not three die, not four, two.

That's like examining the lone 128-bit memory controller on M1 and trying to predict the bandwidth of M1 Pro/Max. Similar to Intel and AMD, Apple has likely left out features on the Pro/Max silicon to save space.

The codename prediction has been correct for 3 out of 4 products. But somehow, the 4th product is way off and it's not a Jade? Seems extremely unlikely to me.

You were right to say that Apple has to decide on priorities. I'm not sure why you aren't considering the possibility that one such priority call was simply not addressing the Mac Pro product segment in the M1 family of chips.

Consider that Apple announced the transition would take 2 years. It's clear they've taken a bottom-up approach, so that puts the Mac Pro last. If their planners knew it would ship 2 full calendar years after the first M1 product launch, why not try to at least make the first-gen Mac Pro silicon a member of the M2/A15 technology generation, if not the M3/A16? It would look slightly silly if the brand new Mac Pro chip was 2 generations behind in single thread CPU performance.

In the above, I am assuming they plan to make a M series derivative of A15. But even that's not a given. Another priority call they may have made is to put Macs on a 2 year silicon cadence instead of 1 year like phones. They already were doing a pattern like that with high end tablet chips - iPad Pro went from A10X to A12X/Z to M1 (which would've been called A14X if not for putting Apple Silicon in the Mac). There was no A11X or A13X.

Which looks sillier? A high-risk workstation chip based on A15, but with the potential of being delayed by 6 months. Or on-track and on-time A14 derivative which doesn't take advantage of the +10% IPC of A15? Apple isn't fighting for their livelihood here, why take the risk?

Designing these chips is an expensive, high risk project. In all likelihood, Apple will take what they know and build around that. Is it possible they'll take A15 and build a huge workstation chip around that? Sure. This is Apple's first in-house workstation chip. Is it likely they'll go high risk? I say no. Was the M1 launch paired with mini LED and a brand new MacBook chassis?

 

[Ethosik]

I wonder if the Mac Pro will be a bit more unique than the M1+ ones. Will it have less single-core performance for insane multi-core?

 

[JPack]

Okay, 14 cores (with HT) for Alder Lake, 20 for the Ultra. And that doesn't make any difference whether it's a workstation or not.

Mac Pro launched in 2019 can go up to 28 cores based on Xeon W.

Alder Lake maxes out at 16 cores (8 of which E-cores). Today, AMD launched Threadripper Pro for workstations with up to 64 cores.

If you still think Alder Lake is a true workstation processor, I'm not sure what else to tell you.

 

[Juuro]

I wonder if the Mac Pro will be a bit more unique than the M1+ ones. Will it have less single-core performance for insane multi-core?

That sounds like a completely new core and SoC architecture for an extremely niche product.

 

[crazy dave]

That's like examining the lone 128-bit memory controller on M1 and trying to predict the bandwidth of M1 Pro/Max. Similar to Intel and AMD, Apple has likely left out features on the Pro/Max silicon to save space.

Which would require new dies ... like going from the M1 to the Pro/Max.

 

[mr_roboto]

That's like examining the lone 128-bit memory controller on M1 and trying to predict the bandwidth of M1 Pro/Max. Similar to Intel and AMD, Apple has likely left out features on the Pro/Max silicon to save space.

During the video released today John Ternus said "We're adding one last chip to the M1 Family". I don't see any wiggle room there - modern Apple is quite careful when writing the scripts for these things. They leave things ambiguous when they need them to be ambiguous for future reveals, but if they say something definite like this, I think we can trust it. If they had more M1 family tapeouts in the works, they wouldn't have had Ternus speak those words.

Also, an interrupt controller is a tiny, tiny thing. The area difference between one suitable for up to two die and one suitable for up to four should be nothing important. The parsimonious explanation is that they designed this gen's interrupt controller for two die because that's all they planned in this generation.

 

[Tyler O'Bannon]

M2 will launch in fall in a colorful MBA.

By end of year, Mac Pro will follow. It will be 4 and 8 die configs.

4 die config: M2 Ultra Giga

8 die config: M2 Super Mega Ultra Giga

 

[JPack]

During the video released today John Ternus said "We're adding one last chip to the M1 Family". I don't see any wiggle room there - modern Apple is quite careful when writing the scripts for these things. They leave things ambiguous when they need them to be ambiguous for future reveals, but if they say something definite like this, I think we can trust it. If they had more M1 family tapeouts in the works, they wouldn't have had Ternus speak those words.

Also, an interrupt controller is a tiny, tiny thing. The area difference between one suitable for up to two die and one suitable for up to four should be nothing important. The parsimonious explanation is that they designed this gen's interrupt controller for two die because that's all they planned in this generation.

What is the definition of "M1 family"? Why didn't they say M1 architecture? Must derivatives of M1 be called M1-something? Can Apple choose to use W for workstation?

If Apple can call the S5 chip on Watch even though it's identical to S4, I think there's a lot of "wiggle room."

 

[JPack]

Which would require new dies ... like going from the M1 to the Pro/Max.

I don't dispute that. Would it still be based on M1? I would say so, just as M1 Pro is a new die but is based on M1.