Contribute
Register

Mac Studio GPU performance

Status
Not open for further replies.
Hello
Since monday I've been working on Macstudio (M1 Max with 24 GPU cores) at work. I use DaVinci Resolve
I run Geekbench and get a Metal Score of 60000
I was wondering if the score would scale up with the number of GPU cores ?
Do you know Metal Scores for GPUs with 32/48/64 cores ?

My RX6800 XT has a Metal Score of 180000. 3x more :eek:

Food for thought...

Have you assessed how well does 3x benchmark gain translate into actual Resolve perf?

Does some op go 3x faster?

However much faster some op goes, does it translate into a meaningful uptick in your productivity?

Is this enough to justify hack?

A side from that what is the Future of Apple Mx GPU compared to the increassing raw compute power of AMD /NVIDIA GPUs.
I can't imagine how they gonna close the gap

For Apple, there is no gap. They design total systems vertically to achieve overall mix of product goals in consumer space which they single-handedly define.

To compare, it must be Apples to Apples —yarg I slay me...

For example, Apple doesn't choose GPU marketing based on a light-show and bragging rights. They study graphics workloads that typify market segments and tune their offerings to perform effectively in those segments, while consistently advancing generational performance, maintaining a sane schedule of obsolescence, and surprising their customers with advancements and treats along the way.

Today the trick from Apple is to add hardware neural engine and video encoder/decoder to run Davinci quicker, but I'm not sure it's a good path.

The neural/media engines is a case in point: these are a modest cost in HW that's been carefully optimized to add SW value, from camera enhancement and photo library features, to allowing clever tricks like in-line OCR in messages and publishing, to AV signal processing in teleconferencing, etc. Apple can deliver a more effective experience with its vertical orientation to a sizable and fashionable market. Consider the creeping effectiveness of the "blue" vs green bubble in messaging, where Apple-to-Apple users enjoy a far smoother and more effective and enjoyable SMS experience, which just so happens to make non-Apple users feel like their experience is stupid by comparison (which it is) while remaining 100% compatible. Is there any technical reason why other phone makers can't offer similar advances? I don't think so. So it's an example of how design makes a big difference.

One last thing
Apple GPU performance per watt seem equivalent to NVIDIA's :
With ARM chips, Apple seems to make tiny and silent computers with low power consumption (Laptop usage) but how they gonna replace the Mac Pro ??

Any Ideas ??
Thanks

Nvidia latest showing off is just pure carnival. That's why the 4080 looks so garish. It's just fat for the sake of phat. Nvidia is focusing on compute elements for autonomous vehicles, robots, etc. Stuff in another world from PC. Intel ARC GPU proves how boring the PC graphics space really is. The fact that AMD sticks around in desktop GPUs is a sign of weakness, not dominance. Sony Playstation has shown once and for all that gaming has a very finite upper bound on required gaming gfx performance, that can be met at a full system end-user cost of 1:1 for a midline PC gaming GPU.

My view of the Mac Pro is that it is now a legacy form-factor informed by the history of NeXT and Pixar. It's sold into a specific professional segment of "workstation" targetting media production, and some science and engineering, with no aspirations of dominance, just to stay involved in these segments. The history of computing advances are technical swords into plough shares. Apple is a Silicon Valley company; it's in their bones. (oooh too many clichés)

The most distinguishing traits of Mac Pro have been ECC RAM and max RAM, which serve the purpose of ensuring repeatable and correct results for work with a stake in accuracy and consistency, e.g., collaborative content production, CAD, biology, etc., in a form factor of performance and options that agrees with business budgets.

The new Mac Pro should be expected to maintain these traits while bringing proven advances of AppleSi to bear in efficiency and packaging. I predict the Mac Pro will never again be exciting to regular Joes. If it continues to exist, it will remain expensive. It's Apple in dialog with certain industry segments, of which there's no longer any crossover with PC enthusiasts.
 
Food for thought...

Have you assessed how well does 3x benchmark gain translate into actual Resolve perf?

Does some op go 3x faster?

However much faster some op goes, does it translate into a meaningful uptick in your productivity?

Is this enough to justify hack?



For Apple, there is no gap. They design total systems vertically to achieve overall mix of product goals in consumer space which they single-handedly define.

To compare, it must be Apples to Apples —yarg I slay me...

For example, Apple doesn't choose GPU marketing based on a light-show and bragging rights. They study graphics workloads that typify market segments and tune their offerings to perform effectively in those segments, while consistently advancing generational performance, maintaining a sane schedule of obsolescence, and surprising their customers with advancements and treats along the way.



The neural/media engines is a case in point: these are a modest cost in HW that's been carefully optimized to add SW value, from camera enhancement and photo library features, to allowing clever tricks like in-line OCR in messages and publishing, to AV signal processing in teleconferencing, etc. Apple can deliver a more effective experience with its vertical orientation to a sizable and fashionable market. Consider the creeping effectiveness of the "blue" vs green bubble in messaging, where Apple-to-Apple users enjoy a far smoother and more effective and enjoyable SMS experience, which just so happens to make non-Apple users feel like their experience is stupid by comparison (which it is) while remaining 100% compatible. Is there any technical reason why other phone makers can't offer similar advances? I don't think so. So it's an example of how design makes a big difference.



Nvidia latest showing off is just pure carnival. That's why the 4080 looks so garish. It's just fat for the sake of phat. Nvidia is focusing on compute elements for autonomous vehicles, robots, etc. Stuff in another world from PC. Intel ARC GPU proves how boring the PC graphics space really is. The fact that AMD sticks around in desktop GPUs is a sign of weakness, not dominance. Sony Playstation has shown once and for all that gaming has a very finite upper bound on required gaming gfx performance, that can be met at a full system end-user cost of 1:1 for a midline PC gaming GPU.

My view of the Mac Pro is that it is now a legacy form-factor informed by the history of NeXT and Pixar. It's sold into a specific professional segment of "workstation" targetting media production, and some science and engineering, with no aspirations of dominance, just to stay involved in these segments. The history of computing advances are technical swords into plough shares. Apple is a Silicon Valley company; it's in their bones. (oooh too many clichés)

The most distinguishing traits of Mac Pro have been ECC RAM and max RAM, which serve the purpose of ensuring repeatable and correct results for work with a stake in accuracy and consistency, e.g., collaborative content production, CAD, biology, etc., in a form factor of performance and options that agrees with business budgets.

The new Mac Pro should be expected to maintain these traits while bringing proven advances of AppleSi to bear in efficiency and packaging. I predict the Mac Pro will never again be exciting to regular Joes. If it continues to exist, it will remain expensive. It's Apple in dialog with certain industry segments, of which there's no longer any crossover with PC enthusiasts.

ECC will be interesting. If the SoC that Apple uses in the next Mac Pro has ECC and they stick to the shared memory layout, it will probably never be used in anything but Mac Pros. Producing such a low volume SoC will probably be expensive too.

The other thing that I've been wondering about the next Mac Pro is whether or not they will even have PCI-e slots. Since there are no drivers for any AMD or Nvidia video cards in Apple Silicon versions of macOS what will people use the PCI-e slots for if they are there?

For audio and video interface, makers of those products already have very capable Thunderbolt solutions. Would they even bother to produce new Apple Silicon compatible cards if PCI-e slots are only available on such a low volume product?
 
My sense is there's no there there: The Mac Pro died in 2013.

The 2019 was a goodwill gesture Apple could afford to make while it readied a new M1 cube—that just so happens to make the old Xeon's look dumb. Apple was well aware of its own plans when the 2019 was released. So they tarted it up to make a big show, knowing that the new cubes would eclipse the line. Making the 2019 obnoxiously expensive actually helped reception of M1s by making them seem like bargains while holding line on segment Apple PR just had to nurse old expectations along far enough to make the crossing, after which, as we can now see, the Mac market press no longer cares.

In sheer oompf that used to be hallmark of fat Xeon 10 years ago, AMD is out-coring those who care about cores per package (not many content makers, these rigs are servers), while PCs already owned the server farms that pipeline for the media studios. The movie industry doesn't think "Oh if only we could get 32 more cores on the desktop we could finally make Bug's Life Pt. VII". Seeing 7000 EPYC in geekbench just confirms the rough scaling tariffs tradeoffs in symmetric multiprocessing that have been well understood for 25 years. These product roadmaps are old news.

Apple shows us indirectly that today's advances in personal computing are defined by what you choose to do with the kit, where systems now progresses as fast as you can figure it out. In the sense of content authors' expectations—the beating heart of Mac demographic—content creation is a solved space. No one is limited by lack of affordable kit. M1 Studio ended a whole tech track: just unbox and start working.

Apple will produce something with the Pro name, but it's their circus to run the rings any way they want, dancing bears, trapeze thrills, goofy-clown cars...
 
I think Apple is still at the "watch this space" point on their ARM performance curve. They are learning as they develop and as we all know, the Mac Studio - for its size, power consumption and lack of noise - is a stunning achievement. They will build on this because they've taken the old paradigm and changed it. What excites me is to think that there are probably new systems working now on a bench somewhere secret that are even better by many times over.

It seems talk of "cores" is still an attraction in much the same way as a digital camera's mega-pixel size is, sadly, still a selling point. But most of us now know that more pixels doesn't mean better photographs.
 
My sense is there's no there there: The Mac Pro died in 2013.

The 2019 was a goodwill gesture Apple could afford to make while it readied a new M1 cube—that just so happens to make the old Xeon's look dumb. Apple was well aware of its own plans when the 2019 was released. So they tarted it up to make a big show, knowing that the new cubes would eclipse the line. Making the 2019 obnoxiously expensive actually helped reception of M1s by making them seem like bargains while holding line on segment Apple PR just had to nurse old expectations along far enough to make the crossing, after which, as we can now see, the Mac market press no longer cares.

In sheer oompf that used to be hallmark of fat Xeon 10 years ago, AMD is out-coring those who care about cores per package (not many content makers, these rigs are servers), while PCs already owned the server farms that pipeline for the media studios. The movie industry doesn't think "Oh if only we could get 32 more cores on the desktop we could finally make Bug's Life Pt. VII". Seeing 7000 EPYC in geekbench just confirms the rough scaling tariffs tradeoffs in symmetric multiprocessing that have been well understood for 25 years. These product roadmaps are old news.

Apple shows us indirectly that today's advances in personal computing are defined by what you choose to do with the kit, where systems now progresses as fast as you can figure it out. In the sense of content authors' expectations—the beating heart of Mac demographic—content creation is a solved space. No one is limited by lack of affordable kit. M1 Studio ended a whole tech track: just unbox and start working.

Apple will produce something with the Pro name, but it's their circus to run the rings any way they want, dancing bears, trapeze thrills, goofy-clown cars...

Yeah. I think the MacPro7,1 was made to appease a very small, niche clientele and critics. All the ones who were hoping to see a modernized MacPro5,1 no longer had anything to complain about. I really wish we knew how many of those MacPro7,1s were sold.

The majority of my friends IRL are Mac users. I have yet to hear any of them complain about Apple Silicon nor has any of them called with any technical questions. They are not the type to go looking for benchmarks to compare how their MacBooks perform compared to the latest Ryzen or RTX GPU. They are only worried about whether everything is working smoothly for them and it has. I think Apple has a raging success on their hands.

As Intel support gets phased out there will be even fewer apples to apples comparisons. As I've said before, I don't care how powerful or fast the latest Intel or AMD offerings are, if it can't run macOS it's meaningless to me.

I think that Apple has done a lot for the Arm CPUs in general. I see power efficiency as part of conversations far more than I ever had before. I also see far more interest in the use of Arm systems among enthusiasts. I think there's definitely a shift in the consumers and that's a good thing.
 
thanks for all yours thoughts
very instructive
 
ECC will be interesting. If the SoC that Apple uses in the next Mac Pro has ECC and they stick to the shared memory layout, it will probably never be used in anything but Mac Pros. Producing such a low volume SoC will probably be expensive too.
Irritating alternative: All AppleSilicon SoC can do ECC, but Apple does not solder the additional chips for parity on "consumer" machines and castrate all but the MacPro…

The other thing that I've been wondering about the next Mac Pro is whether or not they will even have PCI-e slots. Since there are no drivers for any AMD or Nvidia video cards in Apple Silicon versions of macOS what will people use the PCI-e slots for if they are there?
Chicken-and-egg situation here! If no AppleSilicon Mac has a PCIe slot, no-one will write drivers for the ARM version of macOS. As soon as Apple does release a Mac with PCIe slots, card vendors will step in.
Drivers do not have to come from Apple, and PCIe slots are not only for dGPU.

For audio and video interface, makers of those products already have very capable Thunderbolt solutions. Would they even bother to produce new Apple Silicon compatible cards if PCI-e slots are only available on such a low volume product?
"Thunderbolt solutions" would be expected to rely on Thunderbolt embedding 4 lanes of PCIe 3.0. The device remains a PCIe device, and needs a driver—the same driver, regardless whether it is a card in an internal PCIe slot, a card in an external enclosure, or a custom PCB in a device in a Thunderbolt chain.
For high-end and low-volume solutions, I could imagine that "the market" is small and largely overlaps with "MacPro owners" (or "guys who look at a $50k MacPro and say: Only? I'll take ten. For a start.").

Also PCIe 3.0 x4 is nothing to brag about these days. It's the same bandwidth as a single PCIe 5.0 lane. Even a hypothetical update to PCIe 4.0 x4 in "Thunderbolt 5" will fail to impress.
 
Irritating alternative: All AppleSilicon SoC can do ECC, but Apple does not solder the additional chips for parity on "consumer" machines and castrate all but the MacPro…


Chicken-and-egg situation here! If no AppleSilicon Mac has a PCIe slot, no-one will write drivers for the ARM version of macOS. As soon as Apple does release a Mac with PCIe slots, card vendors will step in.
Drivers do not have to come from Apple, and PCIe slots are not only for dGPU.


"Thunderbolt solutions" would be expected to rely on Thunderbolt embedding 4 lanes of PCIe 3.0. The device remains a PCIe device, and needs a driver—the same driver, regardless whether it is a card in an internal PCIe slot, a card in an external enclosure, or a custom PCB in a device in a Thunderbolt chain.
For high-end and low-volume solutions, I could imagine that "the market" is small and largely overlaps with "MacPro owners" (or "guys who look at a $50k MacPro and say: Only? I'll take ten. For a start.").

Also PCIe 3.0 x4 is nothing to brag about these days. It's the same bandwidth as a single PCIe 5.0 lane. Even a hypothetical update to PCIe 4.0 x4 in "Thunderbolt 5" will fail to impress.

Yes. Chicken and egg. But if you were a hardware maker, would you waste development efforts on a PCI-e device for the handful of Mac Pros out there or just make a Thunderbolt device that every single Apple Silicon Mac has?

Yes, 4 lanes. But that's plenty for the audio and video interface that are out there today. Apollo, Presonus, Motu, Blackmagic. They've all embraced Thunderbolt and the devices work fine.

It has long been Apple's desire to kill off PCI-e slots in Macs and this is their best opportunity. What better time to kill off the slots than when there are no compatible cards and no real practical need for them? Remember, between 2014-2019, Apple didn't sell any Macs with PCI-e slots.

The things that use the most bandwidth are NVMe storage and GPUs. Apple obviously doesn't want to give us easily upgradable storage or they would put M.2 slots in all the Apple Silicon Macs. There are no drivers for GPUs. So what incentive does Apple have in putting PCI-e 5.0 slots in?

RAID card? Look at the MacPro7,1. Even if you got an internal RAID card, where would you mount the drives?

When looking at the MacPro7,1, the primary uses for the PCI-e slots were GPUs, Afterburner cards, audio/video interface cards. There are no Apple Silicon compatible video cards... Afterburner is useless since the built-in Media Engines on Apple Silicon are far better... AV interfaces are all available via Thunderbolt.
 
Last edited:
Remember, between 2014-2019, Apple didn't sell any Macs with PCI-e slots.
…and users complained so much that Apple eventually reintroduced a tower MacPro with internal expansion slots.

I do expect that the tower form factor will live on, and for the sake of flexibility I hope that there will be standard expansion slots and connectors in there.

The things that use the most bandwidth are NVMe storage and GPUs.
Both excellent reasons. Those who put one or two dual W6800 MPX modules in their MacPro7,1 will likely not be satisfied by any Apple SoC.
Fast, local and user installable storage certainly has a broader appeal. Anything would do, but in this class of system I would expect U.2/U.3, or even E.3, connectors rather than M.2. Otherwise, PCIe slots would do, especially if bifurcation is possible. (I chuckled at the mention of a "RAID card"—who wants spinning rust in a desktop?—but a tri-mode LSI 9400 controller could be a real option.)

There's also networking. Base-T and copper cable are a technological dead end. And if power users are expected to deport external storage and/or GPU work to servers, they'll want better networking than 10 GbE on copper. So a slot for 40 GbE (QSFP+) or 25 GbE (SFP28) would be nice.
You can put that network card in a Thunderbolt enclosure, like external storage, but each card or each drive will basically want the whole port bandwidth; daisy-chaining devices will be a bottleneck and there will never be enough ports for everything.

Plus whatever specialised card some obscure vendor may come up with… Flexibility.
 
…and users complained so much that Apple eventually reintroduced a tower MacPro with internal expansion slots.

I do expect that the tower form factor will live on, and for the sake of flexibility I hope that there will be standard expansion slots and connectors in there.


Both excellent reasons. Those who put one or two dual W6800 MPX modules in their MacPro7,1 will likely not be satisfied by any Apple SoC.
Fast, local and user installable storage certainly has a broader appeal. Anything would do, but in this class of system I would expect U.2/U.3, or even E.3, connectors rather than M.2. Otherwise, PCIe slots would do, especially if bifurcation is possible. (I chuckled at the mention of a "RAID card"—who wants spinning rust in a desktop?—but a tri-mode LSI 9400 controller could be a real option.)

There's also networking. Base-T and copper cable are a technological dead end. And if power users are expected to deport external storage and/or GPU work to servers, they'll want better networking than 10 GbE on copper. So a slot for 40 GbE (QSFP+) or 25 GbE (SFP28) would be nice.
You can put that network card in a Thunderbolt enclosure, like external storage, but each card or each drive will basically want the whole port bandwidth; daisy-chaining devices will be a bottleneck and there will never be enough ports for everything.

Plus whatever specialised card some obscure vendor may come up with… Flexibility.

At the time they complained, there were compatible PCI-e card that they can use. What PCI-e cards will you plug in an Apple Silicon Mac?

Yes, it's great for storage. But if Apple wanted everyone to be able to upgrade SSD storage easily, they would've used M.2 NVMe SSDs on all the Apple Silicon Macs.

What eGPUs are Apple Silicon Mac users using?

When the MacPro7,1 was introduced, I didn't see Nvidia rushing to release new web drivers. There just wasn't enough Macs with PCI-e slots to make it worth their while.

Yes, faster NICs are nice if anyone makes drivers for them. Who knows, Apple may make faster NICs options like what they did with the Mac minis.

The M1 Max and M1 Ultra have enough Thunderbolt channels for users to setup a router if they wanted there's no need to daisy chain.
 
Status
Not open for further replies.
Back
Top