Thoughts on the Mac Studio & Mac Pro

With these machines starting to ship and the embargoes dropping on the review units I’m starting to get a clearer picture of what the product is. In many ways this is the true successor to the trash can Mac Pro. Very powerful, specifically aimed at the new jobs around media (audio and video) for both the influencer YouTube crowd as well as a significant chunk of the TV and film production market. 18 parallel streams of 8K ProRes video is crazy stuff.

There’s nothing in any of my current work that requires this kind of power. I’m perfectly suited to the MacBook Pro with a “lowly” M1 Pro and 32GB of RAM. Given that there is practically no downside to living the desktop-laptop lifestyle with the M1 Pro and M1 Max, I’ll be staying here for the foreseeable future, hoping that I’ll be doing a little more travel than in recent times.

With that said, I’m seeing the reviews with the usual complaints around Apple systems regarding the crazy margins for the upgrades to RAM and storage. Starting with the RAM we are stuck as the new design of the SoC means that it’s really on the SoC itself, not just some regular DDR memory in a small form factor that is soldered to a board. So this is one that we’re just going to have to live with. On the upside, having the memory on the processor die gets us blazing fast memory I/O to and from the processor, GPU cores and the neural engines, so it’s not a complete loss. Like everything in systems design it’s a tradeoff and in the grand scheme of things a pretty reasonable one.

On the storage front, I think that people complaining about the pricing of the onboard storage haven’t thought through the options. First off, we’re talking about a desktop so having a permanently attached piece of external storage is not the same kind of issue as an external drive with a laptop. Plus with Thunderbolt 4 we have lots of low latency bandwidth options available.

If you need high performance storage, you can use something like the OWC Express 4M2 which holds 4 NVMe M.2 cards. Their marketing material says up to 16TB, but there are now 8TB M.2 NVMe drives on the market (Sabrent, Corsair) so this should go up to 32TB now. And they are daisy-chainable, although you will run into the 40Gbps limit if you parallelise the I/O between them so you’ll want to spread them across different ports on the Studio to get the highest possible performance.

OWC also has their Thunderbay Flex 8 which get you all of the expansion you need in what looks like a standalone tower chassis including a PCIe slot plus 8 drive bays that take SATA disks and SSD plus U.2 form factor NVMe SSDs.

Don’t forget that you also have a native 10GbE network card for connecting to remote storage and if you need more, there’s stuff like the ATTO Thunderlink that goes to 25GbE while waiting for a Mac Pro with internal PCIe slots to go higher.

With a desktop computer we can invest in an ecosystem around it that will last through multiple generations of computing hardware. You replace the computer, but keep your monitor, external storage, high end camera, microphones, etc. These days for my desktop computers I take the 512GB internal storage and then move the user home directories to an external NVMe. Migrations are quick as there are is only the system and applications to migrate.

When the computer is replaced, I keep all of the value in the setup around it.

Mac Pro

The teaser at the moment is now that we’ve seen the Mac Studio, what are the options left for the Mac Pro? Specifically if we take the statement that the M1 Ultra is the last processor in the M1 line at face value.

The key differentiators of the Mac Pro from everything else in the lineup is that it is modular in the sense that we can open it up and add industry standard expansion cards and memory as well as having the necessary connectors to add additional storage devices.

Since the release of the M1 and its unified architecture it would seem that Apple is all in on the UMA design and unless there are more secrets hiding in the UltraFusion interconnect, we have hit the CPU ceiling for this design. There are some rumours floating around that it may be possible to stack two M1 Ultra chips and tie them into the same UltraFusion bus but that leaves aside the complexity of the SoC to mother board connections, heat dissipation and so on. I don’t know enough about that level of chip and board design to have an opinion on whether this is possible or even worthwhile or not.

But among the things that I do know about, MacOS does know how to do NUMA based multiprocessing so while it may not be as good as an UMA design for straight line raw performance, running multiple M1 Ultras in parallel is not off the table from a technical standpoint. The OS certainly knows how to do this for CPU loads and with some work could integrate similar load-balancing and allocation systems for the other components on the SoC like the GPUs and Neural Engines. But from the most basic design perspective, this is the same as modern Intel multiprocessor systems where memory is allocated in banks to a given CPU. So you can end up in situations where the CPU needs access to information residing on the memory bank of the other processor and takes a latency hit in traversing the CPU interconnect. But this is manageable. Some very specific workloads won’t get the linear 2x performance that would come from an UMA architecture tying two M1 Ultras together, but nobody else is either.

Then there is the question of memory. Assuming the possibility of a dual processor Mac Pro using two M1 Ultras, we still have a RAM ceiling of 256Gb with the two SoC at their maximum configuration. Which is a decent amount of memory, but there are workloads out there that are very RAM hungry and could still use even more.

Which brings me to something that I do know about and that is the use of Intel Optane as a new tier of memory. Intel is shipping Optane in both a storage and a memory form factor. From the storage end of things it’s really fast, really expensive (on a cost/Gb) storage media that has a very useful place in the storage hierarchy for things like cache where its resilience and low latency IO performance excel. But Optane can also be delivered as a memory DIMM and with compatible motherboards and the right software it can serve as reasonably fast and cheap RAM. As per the usual computing constructs, you can either use it as a tier (in that “cold” data is moved from the more limited expensive DRAM to the Optane, and “hot” data promoted ) or it becomes the main memory and your DRAM becomes a cache front-end to the Optane. These techniques are proven and available in products like MemVerge’s Memory Machine. The Memory Machine also does lots of other astoundingly cool things like memory snapshots at the process level.

I’m not suggesting that Apple would use Optane in a future Mac Pro, but they could very easily apply the concepts of memory tiers to permit the extension of the memory capacity beyond what is on the SoC to access “slow” DRAM that is installed in regular DIMM slots on the motherboard. So it’s not impossible to imagine a future Mac Pro with DIMM slots that add capacity to the fast RAM on the SoC.

That said, I’m not betting on it, but at least we know there are possibilities.