9 ms·

Looks like all the people saying "just start fusing those M1 CPU's into bigger ones" were right, that's basically what they did here (fused two M1 Max'es togeth

Looks like all the people saying "just start fusing those M1 CPU's into bigger ones" were right, that's basically what they did here (fused two M1 Max'es together). And since the presenter mentioned the Mac Pro would come on another day, I wonder if they'll just do 4x M1 Max for that.

ksec 5y ago
>Looks like all the people saying "just start fusing those M1 CPU's into bigger ones" were right, Well they were only correct that Apple managed to hide a whole section of Die Image. ( Which is actually genius ) Otherwise it wouldn't have made any sense. Likely to be using CoWoS from TSMC [1] since the bandwidth numbers fits. But needs further confirmation. [1] https://en.wikichip.org/wiki/tsmc/cowos https://en.wikichip.org/wiki/tsmc/cowos
- kodah 5y ago
  I've been using a Vega-M for some time which I think follows this model. It's really great.
- kalleboo 5y ago
  The interconnect area was already known from third-party die photos of the M1 Max https://twitter.com/vadimyuryev/status/1466526403331952644 https://twitter.com/vadimyuryev/status/1466526403331952644
  - ksec 5y ago
    They were not known when people keep rampling about the the Mark Gurman report on "just start fusing those M1 CPU's into bigger ones". I wrote about it three months ago. https://news.ycombinator.com/item?id=29430817 https://news.ycombinator.com/item?id=29430817
wilg 5y ago
https://hypercritical.co/2021/05/21/images/city-of-chiplets.png https://hypercritical.co/2021/05/21/images/city-of-chiplets....
iSnow 5y ago
>I wonder if they'll just do 4x M1 Max for that. They'll be running out of names for that thing. M1 Ultra II would be lame, so M1 Extreme? M1 Steve?
- tiernano 5y ago
  M1 Max Pro... :P
  - ceejayoz 5y ago
    "iPhone 14 Pro Max, powered by the M1 Max Pro".
- gordon_freeman 5y ago
  M1 Hyper or M1 Ludicrous:)
  - jazzyjackson 5y ago
    M1 Houndstooth
  - bacro 5y ago
    M1 God
- sdenton4 5y ago
  Just need to increment the 1 instead... Eventually moving into using letters instead of numbers, until we end up with the MK-ULTRA chip.
  - Roboprog 5y ago
    I suspect they will have a different naming convention after they get to M4. There might be some hesitance installing an M5. You should stay out of the way if the machine learning core needs more power. I guess by the time they get to M5, anyone old enough to get the reference will have retired.
  - 5y ago
zitterbewegung 5y ago
They might have to have the unified memory more dense to get to 1.5 TB max of RAM on the machine (also since this would be originally shared with a GPU). Maybe they could stack the RAM on the SoC or just get the RAM at a lower process node.
- paulmd 5y ago
  The M1 Max/Ultra is already extremely dense design for that approach, it's really almost as dense as you can make it. There's packages stacked on top, and around, etc. I guess you could put more memory on the backside but that's not going to do more than double it, assuming it even has the pinout for that (let's say you could run it in clamshell mode like GDDR, no idea if that's actually possible, but just hypothetically). The thing is they're at 128GB which is way way far from 1.5TB. You're not going to find a way to get 12x the memory while still doing the embedded memory packages. Maybe I'll be pleasantly surprised but it seems like they're either going to switch to (R/LR)DIMMs for the Mac Pro or else it's going to be a "down" generation. And to be fair that's fine, they'll be making Intel Mac Pros for a while longer (just like with the other product segments), they don't have to have every single metric be better, they can put out something that only does 256GB or 512GB or whatever and that would be fine for a lot of people.
  - my123 5y ago
    > You're not going to find a way to get 12x the memory while still doing the embedded memory packages. https://www.anandtech.com/show/17058/samsung-announces-lpddr5x-at-85gbps https://www.anandtech.com/show/17058/samsung-announces-lpddr... > It’s also possible to allow for 64GB memory modules of a single package, which would correspond to 32 dies. It is possible, and I guess that NVIDIA’s Grace server CPU will use those massive capacity LPDDR5X modules too. The M1 Ultra has 8 memory packages today, and Apple could also use 32-bit wide ones (instead of 64-bit) if they want more chips.
  - MBCook 5y ago
    Do people really need 1.5 TB of unified memory? If you had 128 GB of unified memory and another pool of 1.5 TB or whatever of “slower” memory (more like normal speed on the Intel/AMD side) would that work? You (or the OS or the chip) could page things in and out if the unified memory. Treat unified memory as a MEGA L3 cache. Depending on how it’s done it may not be transparent if you want the best performance. But would it work?
kasperni 5y ago
> I wonder if they'll just do 4x M1 Max for that. Unlikely, M1 Ultra is the last chip in the M1 family according to Apple [1]. "M1 Ultra completes the M1 family as the world’s most powerful and capable chip for a personal computer.”" [1] https://www.apple.com/newsroom/2022/03/apple-unveils-m1-ultra-the-worlds-most-powerful-chip-for-a-personal-computer/ https://www.apple.com/newsroom/2022/03/apple-unveils-m1-ultr...
- radlad 5y ago
  In a previous Apple press release[1] they said: > The Mac is now one year into its two-year transition to Apple silicon, and M1 Pro and M1 Max represent another huge step forward. These are the most powerful and capable chips Apple has ever created, and together with M1, they form a family of chips that lead the industry in performance, custom technologies, and power efficiency. I think it is just as likely that they mean "completes the family [as it stands today]" as they do "completes the family [permanently]." [1] https://www.apple.com/newsroom/2021/10/introducing-m1-pro-and-m1-max-the-most-powerful-chips-apple-has-ever-built/ https://www.apple.com/newsroom/2021/10/introducing-m1-pro-an... edit: This comment around SoC code names is worth a look too: https://news.ycombinator.com/item?id=30605713 https://news.ycombinator.com/item?id=30605713
- MuffinFlavored 5y ago
  What is M2 really going to be difference wise?
  - 1123581321 5y ago
    ~15-20% faster if releases start this year, plus whatever optimizations learned from M1 in wide release such as perhaps tuning the silicon allocation given the various system. If next year, M2 or M3 (get it) will use Taiwan Semi's so-called 3nm, which should be a significant jump just like 7-5nm several years ago for the phones and iPads.
    - masklinn 5y ago
      Hopefully one of the changes of the M2 design will be a better decorrelation of RAM and cores count. They’d need that anyway for a Mac Pro replacement (128GB wouldn’t cut it for everyone), but even for smaller config it’s frustrating being limited to 16G on the M1 and 32 on the Pro. Just because I need more RAM doesn’t mean I want the extra size and heat or whatever.
2OEH8eoCRo0 5y ago
Right. Still riding gains from the node shrink and on package memory. Could AMD/Intel follow suit and package memory as an additional layer of cache? I worry that we are being dazzled by the performance at the cost of more integration and less freedom.
- scns 5y ago
  The next CPU coming from AMD will be the 5800X3D with 96MB cache. They stack 64MB L3 on top. Rumours say it comes out 20th of April. edit: typo + stacking + rumoured date
stjohnswarts 5y ago
You don't just "fuse" two chips together willnilly. that was designed in from the beginning for the architecture for future implementation.
MangoCoffee 5y ago
its a chiplet design. whenever people ask what we going to do after 1nm...well, we can combine two chips into one
kzrdude 5y ago
Throwing more silicon at it, like this, sounds extremely expensive or price-inefficient. It's at least two separate chips combined together. That makes more sense, mitigates the problem.
arvinsim 5y ago
The CPU trend nowadays seem to be combining chiplets and having performance and efficiency cores.
truculent 5y ago
My guess would be an M1 Ultra with additional modular expansions for specific purposes. I.e. instead of having a GPU, you could add a tensor processing module if you need to do machine learning, or a graphics processing chip for video production, and so on.

ksec 5y ago

>Looks like all the people saying "just start fusing those M1 CPU's into bigger ones" were right, Well they were only correct that Apple managed to hide a whole section of Die Image. ( Which is actually genius ) Otherwise it wouldn't have made any sense. Likely to be using CoWoS from TSMC [1] since the bandwidth numbers fits. But needs further confirmation. [1] https://en.wikichip.org/wiki/tsmc/cowos https://en.wikichip.org/wiki/tsmc/cowos

kodah 5y ago
I've been using a Vega-M for some time which I think follows this model. It's really great.
kalleboo 5y ago
The interconnect area was already known from third-party die photos of the M1 Max https://twitter.com/vadimyuryev/status/1466526403331952644 https://twitter.com/vadimyuryev/status/1466526403331952644
- ksec 5y ago
  They were not known when people keep rampling about the the Mark Gurman report on "just start fusing those M1 CPU's into bigger ones". I wrote about it three months ago. https://news.ycombinator.com/item?id=29430817 https://news.ycombinator.com/item?id=29430817

wilg 5y ago

https://hypercritical.co/2021/05/21/images/city-of-chiplets.png https://hypercritical.co/2021/05/21/images/city-of-chiplets....

iSnow 5y ago

>I wonder if they'll just do 4x M1 Max for that. They'll be running out of names for that thing. M1 Ultra II would be lame, so M1 Extreme? M1 Steve?

tiernano 5y ago
M1 Max Pro... :P
- ceejayoz 5y ago
  "iPhone 14 Pro Max, powered by the M1 Max Pro".
gordon_freeman 5y ago
M1 Hyper or M1 Ludicrous:)
- jazzyjackson 5y ago
  M1 Houndstooth
- bacro 5y ago
  M1 God
sdenton4 5y ago
Just need to increment the 1 instead... Eventually moving into using letters instead of numbers, until we end up with the MK-ULTRA chip.
- Roboprog 5y ago
  I suspect they will have a different naming convention after they get to M4. There might be some hesitance installing an M5. You should stay out of the way if the machine learning core needs more power. I guess by the time they get to M5, anyone old enough to get the reference will have retired.
- 5y ago

zitterbewegung 5y ago

They might have to have the unified memory more dense to get to 1.5 TB max of RAM on the machine (also since this would be originally shared with a GPU). Maybe they could stack the RAM on the SoC or just get the RAM at a lower process node.

paulmd 5y ago
The M1 Max/Ultra is already extremely dense design for that approach, it's really almost as dense as you can make it. There's packages stacked on top, and around, etc. I guess you could put more memory on the backside but that's not going to do more than double it, assuming it even has the pinout for that (let's say you could run it in clamshell mode like GDDR, no idea if that's actually possible, but just hypothetically). The thing is they're at 128GB which is way way far from 1.5TB. You're not going to find a way to get 12x the memory while still doing the embedded memory packages. Maybe I'll be pleasantly surprised but it seems like they're either going to switch to (R/LR)DIMMs for the Mac Pro or else it's going to be a "down" generation. And to be fair that's fine, they'll be making Intel Mac Pros for a while longer (just like with the other product segments), they don't have to have every single metric be better, they can put out something that only does 256GB or 512GB or whatever and that would be fine for a lot of people.
- my123 5y ago
  > You're not going to find a way to get 12x the memory while still doing the embedded memory packages. https://www.anandtech.com/show/17058/samsung-announces-lpddr5x-at-85gbps https://www.anandtech.com/show/17058/samsung-announces-lpddr... > It’s also possible to allow for 64GB memory modules of a single package, which would correspond to 32 dies. It is possible, and I guess that NVIDIA’s Grace server CPU will use those massive capacity LPDDR5X modules too. The M1 Ultra has 8 memory packages today, and Apple could also use 32-bit wide ones (instead of 64-bit) if they want more chips.
- MBCook 5y ago
  Do people really need 1.5 TB of unified memory? If you had 128 GB of unified memory and another pool of 1.5 TB or whatever of “slower” memory (more like normal speed on the Intel/AMD side) would that work? You (or the OS or the chip) could page things in and out if the unified memory. Treat unified memory as a MEGA L3 cache. Depending on how it’s done it may not be transparent if you want the best performance. But would it work?

kasperni 5y ago

> I wonder if they'll just do 4x M1 Max for that. Unlikely, M1 Ultra is the last chip in the M1 family according to Apple [1]. "M1 Ultra completes the M1 family as the world’s most powerful and capable chip for a personal computer.”" [1] https://www.apple.com/newsroom/2022/03/apple-unveils-m1-ultra-the-worlds-most-powerful-chip-for-a-personal-computer/ https://www.apple.com/newsroom/2022/03/apple-unveils-m1-ultr...

radlad 5y ago
In a previous Apple press release[1] they said: > The Mac is now one year into its two-year transition to Apple silicon, and M1 Pro and M1 Max represent another huge step forward. These are the most powerful and capable chips Apple has ever created, and together with M1, they form a family of chips that lead the industry in performance, custom technologies, and power efficiency. I think it is just as likely that they mean "completes the family [as it stands today]" as they do "completes the family [permanently]." [1] https://www.apple.com/newsroom/2021/10/introducing-m1-pro-and-m1-max-the-most-powerful-chips-apple-has-ever-built/ https://www.apple.com/newsroom/2021/10/introducing-m1-pro-an... edit: This comment around SoC code names is worth a look too: https://news.ycombinator.com/item?id=30605713 https://news.ycombinator.com/item?id=30605713
MuffinFlavored 5y ago
What is M2 really going to be difference wise?
- 1123581321 5y ago
  ~15-20% faster if releases start this year, plus whatever optimizations learned from M1 in wide release such as perhaps tuning the silicon allocation given the various system. If next year, M2 or M3 (get it) will use Taiwan Semi's so-called 3nm, which should be a significant jump just like 7-5nm several years ago for the phones and iPads.
  - masklinn 5y ago
    Hopefully one of the changes of the M2 design will be a better decorrelation of RAM and cores count. They’d need that anyway for a Mac Pro replacement (128GB wouldn’t cut it for everyone), but even for smaller config it’s frustrating being limited to 16G on the M1 and 32 on the Pro. Just because I need more RAM doesn’t mean I want the extra size and heat or whatever.

2OEH8eoCRo0 5y ago

Right. Still riding gains from the node shrink and on package memory. Could AMD/Intel follow suit and package memory as an additional layer of cache? I worry that we are being dazzled by the performance at the cost of more integration and less freedom.

scns 5y ago
The next CPU coming from AMD will be the 5800X3D with 96MB cache. They stack 64MB L3 on top. Rumours say it comes out 20th of April. edit: typo + stacking + rumoured date

stjohnswarts 5y ago

You don't just "fuse" two chips together willnilly. that was designed in from the beginning for the architecture for future implementation.

MangoCoffee 5y ago

its a chiplet design. whenever people ask what we going to do after 1nm...well, we can combine two chips into one

kzrdude 5y ago

Throwing more silicon at it, like this, sounds extremely expensive or price-inefficient. It's at least two separate chips combined together. That makes more sense, mitigates the problem.

arvinsim 5y ago

The CPU trend nowadays seem to be combining chiplets and having performance and efficiency cores.

truculent 5y ago

My guess would be an M1 Ultra with additional modular expansions for specific purposes. I.e. instead of having a GPU, you could add a tensor processing module if you need to do machine learning, or a graphics processing chip for video production, and so on.