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.
MP All Models How will the apple silicon Mac Pro work with discrete graphics if M chips don't support eGPUs?
- Thread starter 257Loner
- Start date
-
- Tags
- apple silicon mac pro
- Sort by reaction score
Honestly, this has now been discussed in like 6 different threads, lol
It would be helpful if you could reference those specific threads since the poster didn't see them.
I, too, can remember that this has been discussed frequently, but of course I can't remember the URLs of these threads. So perhaps OP can just use the search function himself? Not that anybody outside of Apple knows what the next Mac Pro will look like and what kind of GPU it will have and support.
I believe the consensus of those threads was thread was that it wouldn't make sense to give up the advantages of integrated GPUs and unified memory. In other words, they could probably design such a chip but it would be a stupid idea.
AFAIK the only thing stopping discrete GPUs “working“ on Apple Silicon is the lack of OS support and drivers - not trivial to fix, but feasible. That would lose the performance advantages of unified RAM shared directly between the GPU and CPU and a GPU optimised specifically for Metal - so I don‘t think there’s much demand for hanging a mid-range external GPU off a Mac Studio (with bandwidth limited by Thunderbolt).
Problem is, the 2019 Mac Pro used a Xeon chip with a massive 64 lanes of PCIe which could support multiple, high-end GPUs in full-width 16 lane slots. The M1 Max has 4 TB 4 ports, each using the equivalent of 4 lanes of PCIe internally - but those are pretty much needed for other things. Presumably, the Ultra, being 2 Maxs linked together, has 8 TBs (even though the studio ultra only has 6 external ports) and the rumoured “extreme” would have 16 TB ports - I.e. 64 PCIe lanes. So, assuming that those TB4 ports can be “turned back” into PCIe lanes without completely re-designing the chip, a “Mx Extreme” could support enough PCIe lanes to run a couple of high-end GPUs as well as a 128-core integrated GPU and at least 256GB of on-package RAM - and Apple could “build” it using the same basic die designs used across the Mac range.
The question is, though, would a Mac Pro centred on 3rd party GPUs like that be worth Apple’s while? It wouldn’t really offer anything that an Intel/AMD x86 system didn’t and it’s performance would be determined by the same GPUs you could choose for an x86 system. Youre throwing away the ASI advantages of tight OS/hardware integration, unified RAM - and low power consumption isn’t really a selling point on a full tower workstation (esp. when the GPUs will be pulling enough power to boil a kettle). What Apple Silicon is good at is laptops and small-form-factor systems.
I wouldn't be surprised if support only comes via Apple-developed expansion modules when it comes to RAM and GPU. It would be a way to fulfill "expandability" while keeping a hold on the hardware. Like if they created some form of expansion slot based on the interconnect that's used to fuse the M1 Ultra chip together, which then only connected to apple-made RAM and GPU modules. Maybe an updated afterburner media engine specific for the machine that increases capabilities further beyond the SOC itself.
Then add a couple PCIe slots for other things like USB cards, Video/Audio I/O, NVME, etc.
Of course, I'd love for them to continue supporting AMD GPU use, but I wouldn't be surprised GPU expansion becomes possible only via Apple hardware. Would love to be wrong.
Then add a couple PCIe slots for other things like USB cards, Video/Audio I/O, NVME, etc.
Of course, I'd love for them to continue supporting AMD GPU use, but I wouldn't be surprised GPU expansion becomes possible only via Apple hardware. Would love to be wrong.
The Mac Studio seems to be the zenith of what can be made with an integrated SoC Apple chip. How an Apple Silicon-powered Mac Pro will be faster than the Mac Studio AND the competition while offering RAM or PCI Express expandability beyond the Mac Studio are the big questions.
Can we just take a second to reiterate, that the "performance advantage" of Apple Silicon integrated GPU is in the form of a $10k top of the range Mac Studio being spanked by a sub $1000 off-the-shelf card in a 12 year old "obsolete" computer.
Or, that $10k computer maxing out at 4 displays, whereas the obsolete one can host more than one card capable of driving 6x4k displays.
Not to be that guy but the Mac Pro but if I remember off the top of my head, its 92 Lanes, 32 of which (the MPX slots) have direct CPU access whereas the rest are behind the bridge, as not all PCIe slots are equal. It's still a metric assload.
I was going with https://www.intel.com/content/www/u...ssor-38-5m-cache-2-50-ghz/specifications.html which says "64" - but that might just be the direct processor lanes. Anyway, your figure of "a metric donkey load" is adequate precision for this discussion
It's 64 lanes. 2 x 16x lane via direct ports to the CPU. another 2 x 16x lanes via a PCIE switch fabric that you can allocate lanes to via the PCIE expansion utility.
if they can, I would prefer to see some type of custom external module that can connect to the mac pro, studio or macbookpro via TB.
You do lose some bandwidth but my experience with an egpu on a PC [and the intel i9 MBP] has been good.
I have no expectation of 3rd party GPU's working.
You do lose some bandwidth but my experience with an egpu on a PC [and the intel i9 MBP] has been good.
I have no expectation of 3rd party GPU's working.
"Top-end" M1 Ultra Mac Studio is US$8K, not US$10K...
Why would you hook a GPU up via TB if it is for a Mac Pro, just use the proper PCIe slot and lose zero bandwidth...?
Considering Apple is a company based in the United States of America, and they set initial pricing on the US dollar, it would seem only natural for discussion on Mac pricing (especially on a forum that is also US-based) to be US$-centric; maybe clarify which currency you are referencing in the future...
Right, because a USD$8k computer being spanked by a $1k GPU on the specific graphics metric for which the computer is allegedly optimised, is a completely different situation to how things would be if the computer was USD$10k.
Clearly the important part is the price quibble, and not the whole top of the range machine being utterly demolished by an ~impulse-purchase-priced off-the-shelf component installed in a 10-12 year old computer.
...unless you are running the sort of task for which it actually makes sense to buy a Mac - like running FCPx and/or working with ProRes (in which case the M1 Max/Ultra will smoke anything else) or using other Mac-specific software that's been optimised for Metal (most of those benchmarks come from March when the Ultra had just been announced).... and Apple Silicon will do that in a mobile or near-silent small-form-factor system using a fraction of the power of Intel/AMD.
It has long been the case that - if you want the absolute fastest hardware, the latest GPUs etc. and real customisability to run generic software - you could get something cheaper, more powerful and better tailored to your needs by getting PC hardware. Even the 2019 Mac Pro was only "better" if you ignored the option of AMD processors, Scalable multi-CPU Xeon systems, specialist GPU-based computing systems with a dozen GPU cards etc... and if you buy a Mac for gaming you're just holding it wrong.
The Mac Pro is really a side issue for Apple Silicon - the important aspect is what it can achieve in a genuinely portable format like the MacBook Pro, or a SFF "Desktop Appliance" like the Studio. Intel Macs were turning into PC clones with nicer screens and trackpads (and PCs were catching up on the trackpads) - the most that they could even aspire to in terms of raw performance was being "as good as PCs". Apple Silicon means that Apple can actually produce something different that beats PCs in particular applications.
However, trying to emulate the Xeon-W in a "big box 'o' slots" format probably isn't a good use of the tech.
I wish Marcan42 of Asahi Linux didn't delete his account as he had a tweet explaining that Apple Silicon shares a design similar to the Raspberry Pi where there's a complication with the I/O with certain memory ranges affecting dedicated GPUs, and I think it'd require significant changes to the CPU cache to fix it. If someone has a screenshot or a better memory, please do chime in.
Perhaps this is why the Apple Silicon Mac Pro has been delayed. Even if Apple decides to go-full Apple and sells only proprietary MPX GPUs as upgrades, they still have a significant hurdle to overcome to make dGPUs.
I'm moderately hopeful with Apple cancelling the "Apple Silicon Extreme" that there has to be some value addition to the Mac Pros over the Mac Studios and that has to come with some level of modularity and not the "modularity" used to describe the M1 Studio. PCIe slots would still be useful to a select number of users.
Perhaps this is why the Apple Silicon Mac Pro has been delayed. Even if Apple decides to go-full Apple and sells only proprietary MPX GPUs as upgrades, they still have a significant hurdle to overcome to make dGPUs.
I'm moderately hopeful with Apple cancelling the "Apple Silicon Extreme" that there has to be some value addition to the Mac Pros over the Mac Studios and that has to come with some level of modularity and not the "modularity" used to describe the M1 Studio. PCIe slots would still be useful to a select number of users.
You raise a good point. Since Apple Silicon's forte is performance per watt, Apple might be better off selling smaller computers like laptops and Mac Studios where Apple Silicon is competitive rather than Enterprise-class desktops where Apple Silicon is less competitive.
It looks like you already said last Sunday the same thing I just wrote. I really enjoy your analysis. Thanks for sharing!