Discussion Do you think the M-chip in the Mac Pro will have a soldered CPU ?

[FlameTail]

Apple announced in it's 2020 WWDC that wit will transition all Macs to Apple Silicon within 2 years.
Now it's 2022 and 2 years have almost passed. The rumor mill is suggesting that Apple will delay completing their transition- There are 2 more devices for Apple Silicon to be implemented- the Mac Pro and iMac Pro.And it seems that these 2 Macs will continue to rock Intel Chips for a while. But it is CERTAIN, that eventually they will transfer to Apple Silicon.

The Mac Pro and iMac Pro are targeted towards proffesional creators, and their Upgradeability and Modularity is a major part of their appeal.

However, All Apple Silicon so far are soldered chips. As such there is a possibility that the M1 Max Quadro or M1Z or whatever it's called will end up as a Soldered CPU.

Do you think this might happen ? If it does, will you buy the new Mac Pro ?

 

[Ajay]

Possible if the Mac Pros allow for multiple configurable SoCs. iMac Pro’s - seriously doubt it. Though there has been some talk of an M2 in the works. A bit curious, since an M1 Pro/Max have plenty of performance for the iMac Pro, but logical for a Mac Pro with multiprocessor options.

 

[jpiniero]

Soldered CPU and RAM definitely. Apple could still support a second tier type solution with DIMM slots or CXL memory but I dunno how crazy they want to go with that.

 

[Doug S]

100% certain. Sockets are less reliable and you can't buy Apple Silicon CPUs so it isn't like they have to worry about end user upgrades.

 

[FlameTail]

But then a MAC PRO can cost upto 50 000$.

If the CPU is not replaceable, i think it might be highly unpopular. Who wants to pay $50 000 every 5 years or so, when they want to upgrade their processing power ?

 

[Doug S]

But then a MAC PRO can cost upto 50 000$.

If the CPU is not replaceable, i think it might be highly unpopular. Who wants to pay $50 000 every 5 years or so, when they want to upgrade their processing power ?

People who are buying a $50K Mac Pro are using it to generate many times that in income, so they would rather upgrade the entire thing than only part of it. Does Apple even support upgrading the current Mac Pro? Being able to DIY doesn't matter if Apple doesn't support it, because no one is going to do something unsupported to something they paid $50K for. They might take the risk on a $5K Mac Pro, but if an OS update blows up in your face because you aren't using a Xeon they expected/supported, it is your problem not Apple's.

 

[nicalandia]

But then a MAC PRO can cost upto 50 000$.

If the CPU is not replaceable, i think it might be highly unpopular. Who wants to pay $50 000 every 5 years or so, when they want to upgrade their processing power ?

Build a Threadripper Hackingtosh then

 

[deathBOB]

If the CPU is not replaceable, i think it might be highly unpopular. Who wants to pay $50 000 every 5 years or so, when they want to upgrade their processing power ?

Any business who needs support and a warranty?

 

[StinkyPinky]

An apple silicon Mac Pro will probably have 20 and 40 core options, with I assume a max of 256GB of ram (?) unless they find the bandwidth somewhere else. That's a large downgrade in memory capacity. Good for Imac Pro's not so sure about Mac Pro's.

 

[Doug S]

An apple silicon Mac Pro will probably have 20 and 40 core options, with I assume a max of 256GB of ram (?) unless they find the bandwidth somewhere else. That's a large downgrade in memory capacity. Good for Imac Pro's not so sure about Mac Pro's.

LPDDR5 modules can have up to 32 devices, each up to 16Gb in size. That provides 1 GB per bit of width (i.e. up to 64 GB in a typical smartphone's 64 bit wide LPDDR5) With the M1 Max SoC offering 512 bit wide LPDDR5, in a Mac Pro containing four of them up to 2 TB is possible.

With LPDDR5X 32Gb devices are supported, which would make a 4 TB Mac Pro possible. So no downgrade in DRAM is necessary with the Mac Pro, though I expect you'll need to custom order the larger configurations and they may charge more for that since Apple would have to pay more for those lower volume really large LPDDR5 modules. They may not support anything beyond standard configs until the four SoC version of the Mac Pro is released, which sounds like it may follow a two SoC version by at least a year.

 

[Shmee]

I would never buy a computer I couldn't upgrade myself. I might consider older apple products such as the upgrade-able macbook pros, or possibly a new product if it was upgrade-able and reasonably priced, with good performance. As of now though, I have never owned an apple product.

 

[LightningZ71]

In my opinion, the processor packages will certainly be soldered, including their RAM. I think it highly likely that storage will be upgradable. I think that it is possible that a CXL based dram expansion solution will be available, but unlikely.

 

[repoman27]

LPDDR5 modules can have up to 32 devices, each up to 16Gb in size. That provides 1 GB per bit of width (i.e. up to 64 GB in a typical smartphone's 64 bit wide LPDDR5) With the M1 Max SoC offering 512 bit wide LPDDR5, in a Mac Pro containing four of them up to 2 TB is possible.

With LPDDR5X 32Gb devices are supported, which would make a 4 TB Mac Pro possible. So no downgrade in DRAM is necessary with the Mac Pro, though I expect you'll need to custom order the larger configurations and they may charge more for that since Apple would have to pay more for those lower volume really large LPDDR5 modules. They may not support anything beyond standard configs until the four SoC version of the Mac Pro is released, which sounds like it may follow a two SoC version by at least a year.

The practical limitations are the 512-bit bus (32x16) on the Jade C Die and the maximum LPDDR5 die size of 16 Gbit. Those 16 Gbit dies have interfaces that are either 32-bit (2x16) or 16-bit (2x8) in byte-mode, and going beyond dual-rank with LPDDR5 isn't practical. That means 4 GB per 16-bit channel, or 128 GB total, is the max for the M1 Max. That would work out to 512 GB for an M1 Max Quattro. Beyond that, the RAM region of the physical address space for the M1 Pro/Max is currently limited to 3TiB.

 

[Doug S]

The practical limitations are the 512-bit bus (32x16) on the Jade C Die and the maximum LPDDR5 die size of 16 Gbit. Those 16 Gbit dies have interfaces that are either 32-bit (2x16) or 16-bit (2x8) in byte-mode, and going beyond dual-rank with LPDDR5 isn't practical. That means 4 GB per 16-bit channel, or 128 GB total, is the max for the M1 Max. That would work out to 512 GB for an M1 Max Quattro. Beyond that, the RAM region of the physical address space for the M1 Pro/Max is currently limited to 3TiB.

According to Andrei's article about Samsung's LPDDR5X intro it is possible to have 64GB in a single module w/32 dies using 16Gb dies. Whether that's a standard product that's actually on Samsung's or Micron's price lists is another matter, but Apple has enough buying power they could say "we want to buy this many petabytes of LPDDR5 for iPhone and Mac, but oh by the way you have to make available small quantities of 64GB packages as part of the deal."

https://www.anandtech.com/show/17058/samsung-announces-lpddr5x-at-85gbps

The A15 core added a bunch of physical address bits so that limitation (which I thought was less than 3 TB more like 512 GB?) should only be present with M1, not M2 and subsequent. The only possible reasons for Apple to add all those address bits would be for Mac Pro or building very high memory servers for internal use.

 

[Arkaign]

Apple's Mac Pro line kind of meanders the lines between lumbering, slumbering, and blundering depending on the gen, but in general it seems they iterate a little too lazily to be very compelling outside of certain release window timeframes for anyone beyond Mac diehards with a very intrinsic relationship and investment in software and personnel for their desired workflow.

It's not a recipe that seems to support much expansion for the market, and even with the ease of basically off the shelf upgrades they could have offered any number of times on previous models to keep them refreshed in the interim, it has not been uncommon for them to look almost comical just a year or two after release, and then keep those dusty specs on the sales pages for years between gens.

Now, with Intel, AMD, and Nvidia triple stacking GPU competition and the resurgence of x86 wars with Alder+ vs Zen4/EPYC/etc, I don't honestly trust this to change very much in the years ahead.

Let's say they do a big M2* MP in 2023, that they then hibernate with for 3-4 years or more. That's going to just repeat what we've already seen from them. Will they decide that this boutique sector is suddenly worth annual refreshes? Because that's basically the minimum pace to keep competitive in such a market.

And dGPU isn't going anywhere, if anything the gap is widening, not narrowing under any metric comparing high end options.

 

[repoman27]

According to Andrei's article about Samsung's LPDDR5X intro it is possible to have 64GB in a single module w/32 dies using 16Gb dies. Whether that's a standard product that's actually on Samsung's or Micron's price lists is another matter, but Apple has enough buying power they could say "we want to buy this many petabytes of LPDDR5 for iPhone and Mac, but oh by the way you have to make available small quantities of 64GB packages as part of the deal."

https://www.anandtech.com/show/17058/samsung-announces-lpddr5x-at-85gbps

The A15 core added a bunch of physical address bits so that limitation (which I thought was less than 3 TB more like 512 GB?) should only be present with M1, not M2 and subsequent. The only possible reasons for Apple to add all those address bits would be for Mac Pro or building very high memory servers for internal use.

You're assuming that module has a 64-bit interface, which is incredibly unlikely. I'm guessing it's 128-bit, but even still, it would require either quad-ranking or single-channel 16 Gbit dies to hit 64 GB per module. The memory controller on the M1 Pro/Max probably doesn't support LPDDR5X, but either of the previous options would theoretically still be a possibility with LPDDR5. Of course you're also talking about a single package containing four wire-bonded octal-die stacks of SDRAM—which is bananas. And 16 of those packages only gets you to 1 TB for the M1 Max Quattro.

The M1 and previous designs were architecturally limited to 32 GiB of RAM. The M1 Pro/Max increased that to 3 TiB.

Progress Report: October-November 2021 - Asahi Linux

asahilinux.org

 

[Doug S]

You're assuming that module has a 64-bit interface, which is incredibly unlikely. I'm guessing it's 128-bit, but even still, it would require either quad-ranking or single-channel 16 Gbit dies to hit 64 GB per module. The memory controller on the M1 Pro/Max probably doesn't support LPDDR5X, but either of the previous options would theoretically still be a possibility with LPDDR5. Of course you're also talking about a single package containing four wire-bonded octal-die stacks of SDRAM—which is bananas. And 16 of those packages only gets you to 1 TB for the M1 Max Quattro.

The M1 and previous designs were architecturally limited to 32 GiB of RAM. The M1 Pro/Max increased that to 3 TiB.

Progress Report: October-November 2021 - Asahi Linux

asahilinux.org

Thanks for the link. This makes me wonder if the M1 Pro / M1 Max is using the A15 big core...