Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Do you think Apple will release a M1 Mac with a dedicated GPU
- Thread starter RobGwy
- Start date
- Sort by reaction score
Do you see Apple re-architecture Mac OS for a computer with multiple sets of CPU, GPU, neural engine, and many other accelerators?
I don’t see it.
Because it’s really really parallel, almost too parallel, or so parallel that it’s embarrassingly trivial
Wikipedia has a nice section on the term’s originWell, they already have NUMA GPU APIs, the same logic could be extended to the rest of the system. But yeah, it would involve significant hardware and software engineering effort.
I don’t really have any particular intuition here. On one hand, I think that a non-modular Mac Pro could probably be fast enough for most users and it’s definitely simpler to design. On the other hand, Apple has some cool modular tech with MPX and the Mac Pro cooling and it would be odd if they designed all that expensive tech for a machine that is about to enter EOL.
There is a disconnect from reality fallacy there that PCI-e slots are primarily only useful for add-in-cards with discrete GPUs on them. That isn't it. There was a "what is in your slots" thread that existed during the 2014-2018 drought on Mac Pro progress that while a substantial fraction of folks had a "non boot screen" GPU installed in a Mac Pro 2009-2012 they also had :
1. Storage drives ( M.2 drives , 2.5" SSD drives )
2. Network cards (Although more 2-4 1GbE or 1-4 10GbE cards more so than one network socket 40 or 100 GbE card)
3. A bit of mix of above External direct or network attached storage.
4. Audio and/or Video capture cards ( sometimes more that 2-3 range )
If going to push huge data and computation out the "cloud" ( somewhere on the other side of Ethernet connection that is out of immediate sight ; not necessarily 100's - 1,000's of miles away. ) then having faster than Apple SOC network speeds makes a difference.
All really need is a aggregate bandwidth of the add-in-cards to be higher than Thunderbolt 3 (or 4 ... same upper limit) and basically taking a "bandwidth" hit to be on the outside.
Apple pretty much admitted in that April 2017 pow-wow on the Mac Pro revitalization that having one and only one internal drive was not cutting it for a substantial number of their Pro customers. The single 4-8TB SSD get rid of the need to have baseline HDD RAID for entry-midlevel users, but workload capacity and speed requirements were still outstripping single drives.
( the two SSD NAND modules in the iMac Pro and Mac Pro are just a single drive. They are no more "RAID" boosted than any other SSD were reads and writes are about the same speed. All SSDs internally use more than one NAND chip concurrently to hide the huge read/write disparity. Apple's modules are about larger than normal capacity on a small card more so than speed. Apple didn't need have to use super, ultra, bleeding edge density NAND modules to get to 4TB in 2017. They didn't again either in 2019 to move to 8TB. )
Some folks have 20TB active (low latency) data set working sizes and archival targets in the 100's of TB range. A single drive isn't going to cut it there for many, many , many years. Apple doesn't have magic pixie dust to tap dance around that kind of gap because it is primary driven down at the basic physics of implementation that Apple does not do. ( don't make NAND chips any more than they make EUV fabrication tools. )
The cube or retreat to the MP 2013 footprint restraints would a mistake. Apple has the flexibility to take the Mini into that zone. They already have a Mini.
Apple's M-series powered "half sized" Mac Pro doesn't have to support 6-8 PCI-e slots. But if it supports zero, then it will probably have substantive problems getting traction with what is left over of the Mac Pro user base. At least one or two slot support is likely necessary. Maybe those don't have "aux power" connector to drive a top end 3rd party GPU card, but a pretty good chance something will be there ( at least for the audio capture folks and local large scratch space crowd. )
Last edited:
Given the fact that macOS 12 is introducing more low level storage driver API, PCIe slots in an upcoming Mac Pro is almost a given, I think.
WWDC session or new docs pointer? ( I haven't gotten through everything yet).
Depends upon the target coverage though. PCI-e storage could be Thunderbolt on the new DriverKit format.
That's would be weak design path though long term, but there is a path for Apple to cut corners.
Something like RDMA support for very high end Network cards would be a more probable indicator of a PCI-e slot.
It’s not really my area of expertise and I only looked at the diffs. But there are new interfaces for audio driver and custom block storage drivers, abs networking driver seems to have received a large amount of changes as well.
“… Use BlockStorageDeviceDriverKit in conjunction with frameworks like PCIDriverKit to create drivers that can communicate with their hardware using custom storage interconnect protocols. …”
BlockStorageDeviceDriverKit | Apple Developer Documentation
Develop drivers for custom storage devices that communicate with the driver using custom protocols.
like PCIDriverKit isn’t only PCIDriverKit . Pretty good chance could mix this class with the new USBkit . Maybe NetworkKit if Apple isn’t covering iSCSI ( with RDMA ? ) .
not sure how many gyrations SoftRaid and Apple go through but this sounds like a possible useful tool for virtual block devices . Or a virtual machine block storage .
While Apple introduced DriverKit as a future direction for drivers last years . It is somewhat incomplete in covering the entire full scope that IOKit covered . There are holes that Apple probably won’t fill in all the way until macOS 13 ( or 14 if another global disruption pop up ) .
AudioKit changed this year to get rid of need for a “plug-in “ to talk to the driver . That is the kind of iterative refinements that take a couple of dev cycles to finish .
DriverKit isn’t a mature API yet .
That said. It should help a broader ecosystem diversity of plug in PCIe cards , which in turn stoke the pressure to keep at least the Mac Pro as a “ container for cards” features.
There are only GPU MPX cards as far as I know. Multi discrete GPUs support is common amongst operating systems.
But a system that supports multiple SoCs? I just don’t see it. It would only benefit very few people while causing an exponential increase in complexity for developers. In the recent thread about stacking Mac Minis, I guessed that it would take hundreds of millions and years worth of time to make multi-SoC work and work well. Totally not worth it.
I don’t think Apple will support multi SoC. They could make a discrete GPU and support multi GPUs.
As for CPUs, I think a SoC with 64 high performance cores is totally viable on 5nm/3nm. I’m guessing that Apple could even go up to 128 cores on 2nm.
Ampere is shipping a 128 core chip this year.
There is no reason to exponentially complicate the OS with multi SoC support for the benefit of a few people. Just keep adding cores as the node shrinks.
Last edited:
Stacking Minis and having NUMA support via a fast interconnect are two vastly different things however. Such systems exist and programming techniques for using them are well described. You are undoubtedly correct that NUMA would increase the programming difficulty, but then again it will only be used in selected professional applications where utilising more processing power is worth it.
Anyway, as I said before, I don't have strong feelings here. I am merely discussing a technical possibility, not arguing that this is what they should do. But if they want to have modularity, compute boards are the only way I can see they an achieve that. Of course, they could also use a single compute board that you would be able to replace/upgrade down the road.
Now that is extremely unlikely because that would mean completely breaking their GPU programming model. If you think that programming for multiple compute boards and NUMA is challenging, well, programming multiple GPUs some of which have UMA and some of which don't is a much bigger mess.
Depends. Apple might just put 128 cores on an Apple GPU, it would be efficient but it would be tied to Metal like how NVIDIA cards are tied to CUDA, they could also do their own custom dGPU like the "Lifuka" rumor was talking about.
It seems like Apple is not going back to AMD for pro GPU's.
It seems like Apple is not going back to AMD for pro GPU's.
They could switch to headless compute cards (AMD Instinct) that would allow more GP compute to be added. Not sure how that would play with the direction their API's are going in though. Not that I think they would go that route.
The idea with using compute boards is exactly this, but with Apple GPUs of course. I don’t see them using AMD stuff for obvious reasons.
It could be PCIe electrically and signal wise, but not physically. Apple could create a proprietary slot.
They could also make their own bus.
Not quite. A compute board would have its own CPU/GPU/NPU as well as local memory. So applications could do work on such a board taking advantage of all the usual assumptions (unified memory etc.). Communication between different compute boards would of course have to happen over a slower interface (be it PCIe or some sort of Apple proprietary fabric). To take full advantage of such as system applications should be able to check which devices and threads are "local" to each other.
I don't think that's a big question at all to be honest. The CPU/GPU communication will likely work the same way as it work on an M1 today — by all processors being directly connected to the same cache and sharing the memory subsystem.
One can have fast GPUs without using disproportionate amounts of power. Apple GPUs are currently 2-3x more power efficient than Nvidia or AMD.
They will simply scale up their existing architecture and add CPU, GPU and ML cores along with more cache and RAM. They have been crystal clear that their model is SOC with Unified Memory going forward and they have no reason to break that model.
The issue here is that multi gpu setups have been dying for awhile now https://www.techspot.com/news/86956-dual-nvidia-rtx-3090s-sli-neat-but-certainly.html
NVidia has put it on notice for their gpus.
As for pci-slots, I imagine they’ll have some solution that’s very “Apple” in nature. There are plenty of uses for pci cards beyond gpus as has been said, Audio interfaces immediately come to mind.