Why Doesn't the Mac Pro M2 Support Terabyte-Level Memory as Mac Pro 2019?

I suspect that the challenges of creating two-level RAM systems and two-level GPU systems for the M2 Ultra Mac Pro turned out to be more than the designers anticipated. So instead of creating an Apple Silicon analogue of the 2019 Mac Pro, they changed the whole focus of the design yet again.

If you look at the 2013, 2019, and 2023 Mac Pros, each of these represented a very big design departure from its immediate predecessor.

• 2012 aluminum Mac Pro minitower – Standard mini-tower design. Could take up to 128 GB of RAM. Did not have any Thunderbolt support or integration, but took PCIe GPU cards and had very elegant removable drive trays.
• 2013 "trash can" Mac Pro – Equipped with dual GPUs, but not with the ability to accept very beefy replacements. Had six Thunderbolt ports. Could take up to 64 GB of RAM (official limit) or 128 GB of RAM (actual limit). Some users indicated that their applications didn't make good use of the second GPU and that what they needed was the ability to install one very beefy one (with corresponding power and cooling requirements).
• 2019 Intel Mac Pro – Completely redesigned to support very beefy GPUs in a PCIe-compatible and Thunderbolt-3-friendly way (with two MPX bays, each of which contains two PCIe slots and Thunderbolt routing). Could take up to 1.5 TB of RAM.
• 2023 M2 Ultra Mac Pro – Basically a M2 Ultra Mac Studio that has PCIe slots for NON-graphics expansion cards. Cannot take socketed RAM, PCIe graphics cards, or MPX graphics modules. Can have up to 192 GB of RAM that must be purchased at the time you get the machine.

kaz-k wrote:

Hi,

It's because it's a Unified Memory, if you increase more memory cell on the chip, it's getting larger chip die which is getting worse yield product chips, I suppose.

I always assumed that an Apple Silicon Mac Pro desktop would need to have a two-level memory architecture – some "fast" memory arranged around the chip, with the potential for a lot more "slow" memory in DIMM sockets. Then extra caching and/or OS-level-awareness of which memory is "slow" and which is "fast" to try to keep up overall performance.

Otherwise, how would you achieve the same RAM capacity as the 2019 Mac Pro, or anything close to it?

What I didn't consider was that Apple might change the design focus / target market for the machine yet again – just as they did with the (2012 -> 2013) and (2013 -> 2019) changes. There are an awful lot of people for whom 192 GB of RAM is more than enough. I do "get it" that someone in that small niche that requires huge amounts of RAM, and who needs a new computer now, might be forced to buy a non-Apple workstation instead of a M2 Mac Ultra Mac Pro.

There is no Direct replacement.

However, in the around 20 year lifetime of these Mac Pro silver tower machines, boot drive capacity has increased somewhat, but the SPEED of the boot drive has increased from around 50 M Bytes/sec for a fast rotating magnetic drive to in excess of 5,000 M Bytes/sec SSD boot drives currently installed.

We may already have in place the next tier of "memory", beyond what is conventionally though of as 'RAM' memory.

If you set aside some pre-conceived notions, you could create a simulation of Terabyte size today, and it would execute correctly using a Very large Virtual Memory. There might be a lot of Virtual Memory swapped onto the boot drive as swap files. Depending completely on algorithms used, it might (or might not) be slow.

It might take re-thinking some algorithms to get computations to operate on more-local sections, and progress through the entire data set more slowly, to slow the amount of swapping.

But it can be done, today, with ordinary, off-the-shelf hardware.

Like many tough problems, to get really BIG speedups, upgrade the algorithms.

Hi,

It's because it's a Unified Memory, if you increase more memory cell on the chip, it's getting larger chip die which is getting worse yield product chips, I suppose.