I think they’ll be more like 3rd or 4th best on non-Apple Silicon tuned workflows, depending on how well Intel’s GPU’s perform. Fortunately for Apple, the only thing they need to be best at, is running macOS. That, and to ship hardware that performs better than the last iteration of a particular form factor.
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.
Will we ever see M-series chips with higher clock speeds in some systems?
- Thread starter BenRacicot
- Start date
- Sort by reaction score
Intel Macs have supported eGPUs for a while, Apple Silicon never did. Opportunity is always there, not so much if it goes agains company strategy
Actually, I almost expect Apple to take the lead or at least become competitive with Nvidia in raytracing on GPU. When you look at Blender benchmarks, the M2 Max is going toe-to-toe with the desktop RTX 3070 using the CUDA backend (no Optix). That is extremely impressive considering that one is a 40-50W mobile GPU and another is a 200+W desktop with almost 50% higher nominal performance. Once Apple rolls out their raytracing solution, expect a very decent performance boost.
If they can find anyone that wants to pay $3k for a poorer performing solution. Anything tuned for Apple Silicon (where the CPU is essentially writing directly to the GPU’s memory) is going to perform poorly, if at all, on a PCIe bus. Even poorer over an external PCIe bus.
I’m thinking Apple’s going to stick to their current plan. The top end will perform better than the bottom end, but never perform as well as, what at that time may be, a 2000 watt behemoth top end RTX system.
I’m curious about whether the opportunity is there. From the folks at Asahi Linux…
FAQ: Will you be able to use GPUs on the PCIe slots of the new Mac Pro?
Answer: We don't know yet! The PCIe controllers on prior Apple Silicon chips could not support GPU workloads properly, but we have not checked whether this has changed on the M2 Max, which includes an updated PCIe controller. Or, indeed, on the apparently possibly separate PCIe controller block used in the Mac Pro (it's complicated).
We'll keep you up to date with this topic, but right now it could go either way.
I’ll be following this situation as, since Apple doesn’t expect GPU workloads to go external from the SoC, I could imagine that even an updated PCIe controller would still have the same limitation. Which will be disappointing to those folks posting that “Asahi Linux will get other GPU’s working, no problem!” when they’ve said nothing of the sort!
UPDATE: From someone else examining the PCIe situation.
Which may mean that the “features” of their PCIe solution may be close to current (not supporting GPU workloads).
Last edited:
The added cost seems a lot to a hobby user or a gamer. But what if yo are already spending $230K per year to color grade film. So every three to five years you have to spend an extra $3,500 for a new Mac Pro. This does not even att 0.5% to to your cost. Why the high cost per year? The human sitting in that chair in from of the monitor is expensive, and he needs studio space to work in, and he needs medical insurance and vacation time and social security deductions and liability insurance and so on and so on....
The economics of professional equipment is different from the economics of toys and hobby gear. Also, you can take tax deductions from pro gear that is used for your business.
Finally, the only people who buy the PCI slots are those who REALLY need PCI slots. Here is an example. Let's say you have a fiber network for access to digital assets. You own dozens of the cards linked below. Would you buy the Mac Studio or the Mac Pro? There are thousands of people in situations like this.
https://www.avalive.com/Studio-Netw...-Enterprise-FCMP-4G4/118561/productDetail.php
All of the Mac Pros from 2009-2019 had PCI-e switches in them. All of them. The surprise would have been if there was not a PCI-e switch in this new system. All of this commentary about how the sky is falling because there is a switch in a Mac Pro is way , way , way over the top.
And had diddly squat to do whether a linux GPU driver works or not. Two input PCI-e switches are in a reasonable wide variety servers and the sky hasn't fallen down over there. There may be issues up by the root PCI-e controller in the Apple die , but the switch isn't an issue.
There are Mac Pro users who have $30,000 to $60,000 worth of PCIe cards in those slots. Trust me, they see value there.
They’re already competing with AMD’s best in their second duration. I can’t fathom how someone would think they’ll be 3rd or 4th.
Apple's Display controllers aren't really relatively small either. Even if Apple clawed back to what just a GPU 'should' cover they would still likely not get as much into a die as Nvidia does. Apple leans harder on more cache/memory than Nvidia does. Memory is just less denser. Hence , it soaks up more die space.
Buy yes Apple is also dragging around more non GPU stuff. ProRes en/decode competition Apple GPU would smoke Nvidia , but again costs die space that Nvidia will never spend.
Honestly, Apple's designs are currently very well balanced. They may not be the best in all their intended use cases, but they are good enough to offer a good value proposition to their target markets.
Apple has a very good competitive analysis and performance/use case dev teams. So they keep making designs that seem to hit most sweet spots.
The main deficiencies are mainly that their NPU is not the most efficient (in terms of area and power) and they tend to lag behind in terms of connectivity. But those tend to be IPs they have licensed elsewhere.
Their power limits approaches are also not the most adaptive, so they end up leaving lots of performance on the table for their desktop offerings (they don't scale as aggressively as they should from when moving from a laptop to a desktop cooling solution).
The biggest flaw, IMO, is in terms of multi die scalability. They have very poor NUMA behavior when migrating threads from clusters in other dies (for the Ultra parts). And their GPU is even worse in that regard. They also need huge L1 caches for their cores to be performant, which may be an issue when moving to 3nm processes. Since they have seen poor SRAM scaling, so those cores may become even more significant in die space within the SoC.
I think that may be reason why they haven't been able to release a proper replacement for the high end of their product line until now.
But for the majority of use cases, Apple has managed to hit most sweet spots.
Apple has a very good competitive analysis and performance/use case dev teams. So they keep making designs that seem to hit most sweet spots.
The main deficiencies are mainly that their NPU is not the most efficient (in terms of area and power) and they tend to lag behind in terms of connectivity. But those tend to be IPs they have licensed elsewhere.
Their power limits approaches are also not the most adaptive, so they end up leaving lots of performance on the table for their desktop offerings (they don't scale as aggressively as they should from when moving from a laptop to a desktop cooling solution).
The biggest flaw, IMO, is in terms of multi die scalability. They have very poor NUMA behavior when migrating threads from clusters in other dies (for the Ultra parts). And their GPU is even worse in that regard. They also need huge L1 caches for their cores to be performant, which may be an issue when moving to 3nm processes. Since they have seen poor SRAM scaling, so those cores may become even more significant in die space within the SoC.
I think that may be reason why they haven't been able to release a proper replacement for the high end of their product line until now.
But for the majority of use cases, Apple has managed to hit most sweet spots.
Last edited:
Not so, the 2011 imacs had upgradable GPUs, even if not officially supported.
The problem with this approch is that the professional world is often governed more by processes than economics.
To exaggerate a bit, it seems that MBA has become the official state religion in every developed country. Everything is about saving costs and optimizing the processes. If the process says that you can buy a $6k device no questions asked, but you know that a $7k device would be a much better choice, you basically have two options. Either you get the $6k device, or you spend a lot of time and effort to justify the exception to the process.
A small organization is more likely to make the choice based on economics, because a Mac Pro would be a major investment for them. In a large organization governed by processes, prices are more likely binary. Either they are below some arbitrary cutoff, or they aren't.
While I'm a big believer in the notion that there's very little out there that a high-end Mac is superior to a high-end PC for, there are definitely some notable exceptions. Certainly, at least 80% of ProRes based things have to be done on a Mac (ProRes support for Windows DOES exist, but is still limited). You also have latency issues with a lot of high-end audio equipment that are inherent to how Windows handles drivers that you don't have with macOS. So, there are definitely high end use cases where you HAVE to have a Mac with PCIe expansion and where a Thunderbolt breakout box would be insufficient.
Apple has also, more or less identified that 192GB of RAM is sufficient for the vast majority of those use cases. And they did so while the RAM ceiling was still 768GB on 8/12/16 core models and 1.5TB on 24/28 core models.
It is absolutely permanently off the table. They stated in 2020 that Apple Silicon would be built entirely around the unified SoC architecture and that GPUs would exclusively be integrated into the SoC. They reaffirmed that in 2023 both with the release of this Mac Pro and in that interview with John Gruber from last week. You are not getting PCIe graphics cards. You are not getting eGPUs.
Clock speed isn't everything. It's only one factor to performance and, honestly, it's not even that substantial of one.
As others pointed out that was part of Xeon features. These were not mandated by Apple. It just came with it.
In the same way being able to run Windows on an Intel Mac was part of x86 and not a requirement from Apple.
In a 2021 white paper for the Mac Pro it is claimed that Apple pointed to a 192GB RAM typical use case.
Naturally that's 2 years ago but it still was pointed out as a typical use case.
Personally I agree with you. It should have more than 192GB.
Apple's gonna lose sales from this. They know this and it is acceptable to them until 2025 Mac Pro M3 Ultra/Extreme.
I don't think Apple spends a lot of time looking at other folks sweet spots. The analysis seems to be more so confined to the 'rear view mirror' mac products and areas where Apple has had lots of traction. Things that appear largely outside the iPhone/iPad/iPhone market that take off like a rocket .... Apple doesn't have a good well of tracking well on that. ( Siri *cough*. )
that can help or hurt. IF Apple had mindlessly tried to "monkey see , monkey do" Quest/Facebook headset, then they'd be in the same ditch.
Nvidia spends lots more effort folding those into the mix of the GPU cores ( CUDA core is kind of related to Tensor cores like P/E cores are related to Apple's AMX (the core cluster has shared resources for the 'non matrix' core(s) and the 'matrix' core ). However, Tensor cores have a more open interface.(AMX is more of a 'magic box'. )). Apple's approach should allow them to iterate and attack different problems. Apple is far more keen on inference than training. Apple has 1B deployed devices. Nvidia, kind of does not. Apple has AMX also . So about three different ways of 'slicing' the problem with varying degrees of energy efficiency. All going to lead to different die space allocations.
the NPUs can do inferencing with the GPU mostly turned off. In that context, are they still 'not the most efficient'. If bury the NPU cores inside of something else, then to do work that something also has to be mostly awake and running. ( on a lock screen or screen saver mode .. the GPU is mostly hitting the 'snooze bar' and yet need to do face/touch ID. Or track sensors and infer. )
If the power/clocking targets where higher Apple would have even more of an area problem. AMD got to Zen 4c by tossing getting into maximum top fuel dragster, single threaded, pissing match. High clocks waste area. For CPU+GPU dies the very high clocking CPUs cores basically take area from the GPUs. One substantive contributing reason why iGPUs build up a pretty bad reputation.
There is not much of a win in Apple both making the CPU cores bigger and then also not being able to sell them in high numbers to pay for the 'alternative' roll out ( masks , certification/validations , etc. ) .
Both AMD and Intel are detaching decent chunks of there server offerings from that 'top fuel dragster' race also. ( 4c and 'Sierra Forest' ). Once up > 100 CPU cores, it starts to make less and less sense to waste space/core. If have 100 cores letting 99 of them go to sleep is a huge mismatch of workload. Huge amount of expensive die space doing nothing.
This was throughly measured on M2 Ultra already???? M1 had issues even single die. That is why M2 got a substantively upgraded internal mesh network that delivers major uptick in performance with not backhaul Memory upgrade. And the 10,000 pads of Ultra... I have doubts the that UltraFusion really needed much adjustments for the 'new' M2 internal mesh adjustments. ( For two dies in the M1 era that seemed to be overkill. Even the PCI-e additions they made in M2 .. still a bit of overkill. )
Can Apple scale that 4-way. Yeah, there I'm skeptical. Can they do 3-way and that's probably enough.
They have to release something that enough people will pay for. There is lots of opinions on these forums that Apple should chase up into very narrow niche areas . Make a desktop single threaded 'killer' SoC, make a xx9xx GPU 'killer' that make about zero economic sense. It can't be too few that cost way, way too much. Nor does Apple even really have that many resources to chase after them ( barely any Apple 'product' out in those swamps either. )
M3 baseline w/hardware ray-tracing. Remember, Apple’s unlike the other vendors that have to save the best features for the top end. When they ship hardware ray-tracing, it will be performant on the lowest end, and just more performant (with more cores) on the high end. No Extreme required.
So having a central SoC with an UltraFusion interconnect on two opposite sides, and would the SoCs attached on either end of the central SoC have unused UltraFusion interconnects...?
Or maybe the three-way configuration is:
Regular SoC <UltraFusion> GPU SoC <UltraFusion> Regular SoC
The GPU SoC would be the only SoC with two UltraFusion interconnects, with the regular SoCs (Mn Max) having a single UltraFusion interconnect...?
One should assume hardware ray-tracing across the entire M3 range of SoCs, scaling with GPU core count...?
Last edited:
LOL, I guess I should have used an exclamation point rather than a question mark; I mean that I assume hardware ray-tracing will be a part of all Mn SoCs starting with the M3-series...!!!
Where did you find benchmarks comparing the latest Blender version between M2 and Nvidia/AMD?
You can use opendata.blender.org to compare GPUs.
Official blender benchmark data (opendata.blender.org)
It may come to a shock to many but shipments of desktop dGPU have been at a downward slope since as early as 2005. This to me indicates that "perfect" 4090-like performance is not selling all that well.
Use case of decades past are now approaching niche.
The number does not discriminate whether the dGPU was used exclusively for crypto or non-crypto use case.
![9hGBfdHQBWtrbYQKAfFZWD.png]()
Source: https://www.tomshardware.com/news/sales-of-desktop-graphics-cards-hit-20-year-low
Use case of decades past are now approaching niche.
The number does not discriminate whether the dGPU was used exclusively for crypto or non-crypto use case.
Source: https://www.tomshardware.com/news/sales-of-desktop-graphics-cards-hit-20-year-low
Last edited: