Raytracing was released at PhRdm v13 , impossible to forget jokes abuot who(?) was Reyes.
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 40 core ASI with PCI-E slot is around the corner!!
- Thread starter Amethyst
- Start date
- Sort by reaction score
At least soon we will see what actually Apple was doing.
But I dont expect Apple's dGPU (or Compute Modulre whatever) to beat AMD/nVidia, neither having meaningful HW raytracing, last weeks the most consitent rumour about the ASi Mac pro was its getting GPUs from multiple vendors.
I preferred the trash can days. Arguments were far more civilised because it was either the fact it looked like a trash can or was everything no one asked for.
When it comes to workstations, such as the Mac Pro, NOBODY CARES ABOUT WATTAGE. Watch, next you'll start attacking this from a CPU standpoint, and then after that portability, as this always turns out
I have no idea why you are getting so worked up. I listed the specs of these devices, which clearly show that M2 Max is about as fast as Pro Vega II and clearly slower than the 6800 XT.
And again, I don't understand why you are getting so exited when a faster GPU outperforms a slower GPU. By the way, 25% difference in performance is hardly "being smoked". Given the fact that 6800 XT is nominally 40-50% faster, one would expect a larger difference.
By the way, the problem lies clearly with the AMD macOS drivers + Blender optimisations on macOS (or lack thereof). On Windows with AMD's own HIP backend, the 6800 XT scores 2500 points (that's good 60% higher than Metal implementation) and 7900 XTX scores 3600 points. At any rate, if you want to get the full performance out of those GPUs, you won't be using macOS. And on macOS, they are ultimately held back by the driver.
All this data is in the databases you are using.
The result you are quoting is probably the duo W6900X (two GPUs), or maybe some spoofed entry. There are plenty of 7900 XTX (~ 200k-220k on average) and 6900 XT (160k-180k) scores available on GB6 compute.
And yes, 7900 XTX is not nearly as fast as its FTLOPS metric might suggest. It's because the 2x of RNDA3 is mostly a marketing trick to better match Nvidia's specs on paper. RDNA3 has gained the ability to execute two operations per cycle under certain conditions. This allows AMD to claim 2x improvement in peaks FMA rate, but the opportunities for this capability in real applications is limited.
Last edited by a moderator:
If one approach the current state of the mac and look at long time trends it seems to me that on the high end side macs have actually become less and less powerful over the last 10 years compared to general compute trends. And the mac pro is the last hope for a change in that trend. Maybe that’s why a lot of us get emotional. We are leaving a 15 year era of insane economic growth and an “easy “ market where funding and progress almost automatic.
It is not hard to envision a future where people won’t buy another expensive iPhone or mac laptop and where Apple will decline and actually have to fight for its customers again. There are a lot of enthusiasts but if it comes to true needs for sustaining a business, very few will pay a premium for macos i think. For laptop macs are still a golden standard but there is a limit to what people are ready to pay for. Here in Europe, there prices are even more inflated and I mostly see macbbok airs around. Not seen a single desktop mac for years in businesses I have had contact with. There simple is no USP for mac desktops anymore. And there needs to be one. (One could argue that mac minis and mac studios have a USP in being small)
What in the world could be the benefit of a ASi mac pro ?
Will it be faster than a xeon or threadripper system for any task?
Make anything more convenient?
In the end I just find customers like myself left: people that want to use macos because they like it and accept lower performance for the benefit.
I can literally not think of a single task that a mac does so good (in the desktop space) that a PC user would consider switching. Can you?
It is not hard to envision a future where people won’t buy another expensive iPhone or mac laptop and where Apple will decline and actually have to fight for its customers again. There are a lot of enthusiasts but if it comes to true needs for sustaining a business, very few will pay a premium for macos i think. For laptop macs are still a golden standard but there is a limit to what people are ready to pay for. Here in Europe, there prices are even more inflated and I mostly see macbbok airs around. Not seen a single desktop mac for years in businesses I have had contact with. There simple is no USP for mac desktops anymore. And there needs to be one. (One could argue that mac minis and mac studios have a USP in being small)
What in the world could be the benefit of a ASi mac pro ?
Will it be faster than a xeon or threadripper system for any task?
Make anything more convenient?
In the end I just find customers like myself left: people that want to use macos because they like it and accept lower performance for the benefit.
I can literally not think of a single task that a mac does so good (in the desktop space) that a PC user would consider switching. Can you?
You know, all this might sound very impressive to a layman, but I've written enough low-level SIMD code to see how your are constantly switching topics and throwing a bunch of random buzzwords around, but produce very little content. We were talking about raytracing on an outer product processor (like what Apple hardware actually implements). You still failed to provide any example of how this would be done and now you are talking about cryptography?
I feel a little bit like Alice at the tea party trying to make sense out of these posts...
I can totally imagine them opening up the interfaces for hardware virtualisation, which would allow native GPU drivers to operate under virtualised Linux. That would be very useful for a wide range of users. But native Mac support? Very sceptical.
I too think that we will see GPU-only dies at some point, maybe even paired with faster HBM-style RAM. But this is very unlikely to happen in this hardware iteration.
I think you are making some very good points. Not too long time ago, there was a huge performance disparity between desktop and laptop platforms. In the last few years, this gap has shrunk significantly. This is of course affecting the market and the users. It's not that much the case that the high-end got slower, but the base-line got much much faster.
On the Mac side, Apple was ahead in some things as they were pretty much the only company focusing on them, but others have caught up. And since everyone was using the same hardware and increasingly similar designs, there was less and less unique value about the Macs. Hence, the move to Apple Silicon was IMO pretty much inevitable to keep the Mac alive. This was the only way for Apple to offer something different (one way or another) and make the Mac unique again.
We still have very little idea what Apple intends to do about the higher performance market. The Mac Pro was already a tough sell for many years — it was a niche mid-range workstation with decent enough components and ok-ish price-performance ratio, but certainly not something that a normal professional user needs. Apple Silicon does enable new niches and new value for this type of computer, but to get those one would need a significant investment, and it's not clear that Apple is willing to commit to it.
At any rate, I think it's important to remember that Apple is playing a long game. They are perfectly fine with losing a bit of ground today so that they can get sustained growth tomorrow. Their efforts in software space make this very clear. They are heavily investing into Blender for example, even though their hardware is simply not powerful and/or capable enough to compete with Nvidia. At first, it sounds like bad business, but it also means they will have robust software platform in place when their first actually powerful hardware arrives. Similar can be observed in machine learning space. The accelerators we currently have on Macs still come from the mobile phone space. But if you look at the software ecosystem Apple is slowly but surely building, and at the patents, there are good chances that the first desktop-oriented ML accelerator they will ship might surprise a whole bunch of people.
You are free to point out if you see a mistake in what I write or if you disagree with it, and we can discuss it like civilized people. I’ve re-read my posts and was unable to spot any factual errors or substantial flaws in my argumentation. So let’s talk about it. Because so far one of the posters I try to engage with changes the topic any time the conversation be ones more concrete, and another poster starts yelling at me for no apparent reason, and now you saying that I post nonsense apparently for quoting benchmark results and GPU specs. Where is the nonsense?
...pretty much agree with everything you said - my only question is, will there actually be a market for desktop ML?
The two natural homes for ML are probably (a) in the cloud (or private data centre) because that's where all the training data lives, and because it often makes more business sense to rent computing power on demand from the cloud and (b) mobile and embedded devices because that's where the applications of ML that can't rely on fast web access are (biometrics, photography etc. in phones, cars, smart devices...). You're not gonna want to carry a Mac Pro on your back to use the new AR goggles. Both of those are areas where the power consumption & heat vs. performance considerations of Apple Silicon are an advantage. The "traditional" MP is not great as a high-density computing device, and it fits into the niche where the power consumption of the CPU is probably the least important (no batteries, plenty of space for cooling & you're not going to be stuffing hundreds of them into a data centre).
The current Mac Pro is very specifically designed as a standalone, super-powerful (by Mac standards) personal desktop workstation - and, squeezed between the cloud on the one hand and increasingly "good enough" mobile and SFF systems... well, with all due respect to the people who actually do need such a system today, its probably not a market with a long-term future that would be worth developing custom silicon for. Also, as you say, the 7.1 was never that powerful c.f. more specialised PC workstation/server hardware, and what advantage it did have was mostly courtesy of new - at the time - chips from Intel and AMD.
OTOH - even without developing new ML silicon. Apple could probably put 8 Mx Ultras into a 1U rack and use them to build nicely scalable clusters for cloud computing.
Good point. I'd guess that desktop will be used for model development/specialisation, as well as inference. For example, artists or developers using large ML models as assistants could benefit from running the inference locally. Right now it doesn't seem feasible because of how large these models are, but I wouldn't be surprised if they will get more optimised and specialised in the future. Super large model training of course will remain the niche of supercomputers.
P.S. We are just in the beginning of the large ML revolution, who knows what the software and hardware will look like in a couple of years...
Last edited:
I think this is something we can have wide consensus, when apple took the decision to break with Nvidia and open GPU architecture it was specially painful for those using a Mac Pro in research, those people forcefully had to leave the Mac as main tool, and move to Linux systems or someones as me using an Mac just as Linux cluster terminal or editor by remote tools.
Things changed a bit to improve but HPC pro users simple are no More using Mac to run on Mac, few niches improved as machine learning as tensorflow pytorch now have backends on Metal which enable Mac Pro GPUs for quite decent training or inference, even Julia now have a JuliaGPU backend for ASi and metal 3 very good for floating point, even rust support MTX but coding for ASi on rust still Way long to be practical or profitable, still much better landscape than at trashcan times, but people is gone, without a community of developers doing stuff it's quite steep for apple to keep industry pace.
Siri is clear landmark example on how ml sucks on apple, and the next Big thing isn't video at 8k, neither is clear as to bet you home on AR/XR, machine learning, specially Generative Pre trained Transformers (aka GPT) are what drives crazy all minds Now, and sadly the entire Mac/iOS ecosystem isn't up for this, and had to blame Apple decision to cut with Nvidia and kick out HPC developers to favor only video publishers, there's not enough Hollywood YouTube TikTok or porn people to drive Mac development on Long term, the Mac Studio is the evidence you need, those people actually don't need an HPC Mac Pro, Mac studio is enough for them, indeed machine learning drives no meaningful interested on Mac, but is clear apple is aware on this and a remedy is on the way even if apple has to sell it at loss, the next Mac Pro likely to be central on apple ecosystem for AI and XR, so discreet GPU and discreet NPU/TPU cards support should be in the pipeline along a capable Mac Pro, both use cases not particularly requiring 128 or 64 CPU cores, but as much GPU/TPU/NPU cores as you can power with 1600W PSU while easily clustering multiple or by dozens of Mac Pro.
Last edited:
Absolutely, but I'd wager a few bucks on the 'delivery' hardware looking more like an iPhone, Watch or HomePod than a 2019 MacPro.
...meanwhile is Apple even interested in continuing to develop hardware and software for ML development/training and backend...? AFAIK they're not using Macs as the back-end for their current services, whereas they're probably ahead of the game in terms of ML delivery capability in mobile, consumer devices.
Nothing particularly uncivil about those benchmark numbers.
What’s uncivil has been the constant trash talking about the 7,1 machine and the Intel Mac Pro in general by numerous people.
A 7,1 with RX6900XT is a fast machine. Depending on what the new Mac Pro is then the 7,1 will become priced within reach of people to buy second hand or it will retain value. So far even second hand they are fiercely expensive.
What’s uncivil has been the constant trash talking about the 7,1 machine and the Intel Mac Pro in general by numerous people.
A 7,1 with RX6900XT is a fast machine. Depending on what the new Mac Pro is then the 7,1 will become priced within reach of people to buy second hand or it will retain value. So far even second hand they are fiercely expensive.
Totally agree the 7,1 is a great machine. And ya.. if the ASi is a huge dud, I think a lot of 2nd hand 7,1's will be bought. I think the complaints about the 7,1 are mostly about Apple's pricing both on day one, and still as Apple always does, continues to charge full price for antiqued hardware for as long as its sold. I personally had money in hand ready to buy a 7,1, but quickly realized I had to configure it up to $10K to get a usable spec. I wasn't ready to do that on day one. And for the same $6-7K you could build a heck of a threadripper 3960/3970? and RTX 2080ti? combo that would destroy the equivalent spend on the 7,1. I'm sure I would of sunk $10K into a 7,1 at this point if I could of gotten on the treadmill.
People here ranting about mp 7,1 and 8,1 mostly driven by Gurman's rumour (lacking actual cues).
The 7,1 still have long life ahead, not just at updating legacy x86 ecosystem apps, but also as content/development on things not app related as 3D, audio video, even AI/ML as Metal 3 support tensorflow and pytorch which are front-end agnostics as long it's backend is supported and performant, so besides performance doing ML on the Mac Pro 7,1 and 8,1 only will differ on GPU performance (assuming new PCIe GPU won't be back compatible with 7,1).
I have no doubts on apple being capable to deliver an quad m2-max rig, just curious about the selected solution to arrange 4 m2 either linear (like a domino), S overlapped, or my guessed as in an + sharing an octagon-shaped quad bridge or interposer.
We could safe bet it's top CPU configuration will be 4x m2 Max, and max GPU performing below 60tflop fp32 (close to single Rx 7900xtx).
Having concerns on how much and kind of RAM upgradeability I'm convinced technically Apple has no barriers than marketing on providing it with support for discreet RAM modules about its format DIMM, SO-DIMM are the popular choices but Apple don't care on being popular so while I consider the Mac Pro 8,1 RAM to be upgradeable, doesn't means you could buy an barebones Mac Pro and at the same time order at amazing a bunch of cheap ram sticks to evade Apple tax, I doubt Apple to offer only soldered ram with the "unified memory" moto, Pro's (actual pro not gamers) don't buy this bullsh1t as it's just an fancy name for old Shared Memory, also impractical for sales as RAM requirements for Pro workflow vary tremendously, there's no 4 ram sizes fit everyone interested on the Mac Pro.
Across the industry is well know apple to support new non apple GPUs, granted AMD but don't discard Intel's ARC at sometime, even maybe Nvidia but this really requires an miracle.
I can imagine an Rx 7900xtx duo PCIe 5 MPX module, two of these added to quad m2 iGpu performance would peak near 300 Tflop/s shutting mouths as the most powerful sub 1.5kw workstation (also supporting HW accelerated raytracing). Do you think John Ternus will be ashamed?
But I think it's necessary another Big "one more thing...'
Apple own specific ASi GPU or compute card, most GPU are really fast doing general highly parallel calculations on float or integers but are quite deficient at Tensors, an pure ASi TPU/NPU compute module would obliterate even the new Nvidia Hopper H100 on where Nvidia is weak: as on tensor processor' IP Nvidia actually don't have an edge, even Apple may have outsourced it to Jim Keller and going under the radar.
Now imagine Jhon Ternus showing that 'one more thing...' a single 200W TPU/NPU accelerator capable to dethrone Nvidia's still to release H100 GPU at machine learning training?
Hope that ASi TPU NPU not happening outside Dreamland otherwise by first time we will see rows at Apple stores on people wanting an ASi Mac Pro, but not everyone but most AI/ML CEO/CTO would be the ones linings.
The 7,1 still have long life ahead, not just at updating legacy x86 ecosystem apps, but also as content/development on things not app related as 3D, audio video, even AI/ML as Metal 3 support tensorflow and pytorch which are front-end agnostics as long it's backend is supported and performant, so besides performance doing ML on the Mac Pro 7,1 and 8,1 only will differ on GPU performance (assuming new PCIe GPU won't be back compatible with 7,1).
I have no doubts on apple being capable to deliver an quad m2-max rig, just curious about the selected solution to arrange 4 m2 either linear (like a domino), S overlapped, or my guessed as in an + sharing an octagon-shaped quad bridge or interposer.
We could safe bet it's top CPU configuration will be 4x m2 Max, and max GPU performing below 60tflop fp32 (close to single Rx 7900xtx).
Having concerns on how much and kind of RAM upgradeability I'm convinced technically Apple has no barriers than marketing on providing it with support for discreet RAM modules about its format DIMM, SO-DIMM are the popular choices but Apple don't care on being popular so while I consider the Mac Pro 8,1 RAM to be upgradeable, doesn't means you could buy an barebones Mac Pro and at the same time order at amazing a bunch of cheap ram sticks to evade Apple tax, I doubt Apple to offer only soldered ram with the "unified memory" moto, Pro's (actual pro not gamers) don't buy this bullsh1t as it's just an fancy name for old Shared Memory, also impractical for sales as RAM requirements for Pro workflow vary tremendously, there's no 4 ram sizes fit everyone interested on the Mac Pro.
Across the industry is well know apple to support new non apple GPUs, granted AMD but don't discard Intel's ARC at sometime, even maybe Nvidia but this really requires an miracle.
I can imagine an Rx 7900xtx duo PCIe 5 MPX module, two of these added to quad m2 iGpu performance would peak near 300 Tflop/s shutting mouths as the most powerful sub 1.5kw workstation (also supporting HW accelerated raytracing). Do you think John Ternus will be ashamed?
But I think it's necessary another Big "one more thing...'
Apple own specific ASi GPU or compute card, most GPU are really fast doing general highly parallel calculations on float or integers but are quite deficient at Tensors, an pure ASi TPU/NPU compute module would obliterate even the new Nvidia Hopper H100 on where Nvidia is weak: as on tensor processor' IP Nvidia actually don't have an edge, even Apple may have outsourced it to Jim Keller and going under the radar.
Now imagine Jhon Ternus showing that 'one more thing...' a single 200W TPU/NPU accelerator capable to dethrone Nvidia's still to release H100 GPU at machine learning training?
Hope that ASi TPU NPU not happening outside Dreamland otherwise by first time we will see rows at Apple stores on people wanting an ASi Mac Pro, but not everyone but most AI/ML CEO/CTO would be the ones linings.
Last edited:
Will it be a 8,1 though, or will it be a 14,xx...?
Would love to see a quad SoC configuration, just so that & PCIe slots can stand out over a top-end Mac Studio; maybe that will quell the "Apple should just re-name the Mac Studio the Mac Pro" folks......
To be fair, is not Apple's UMA a bit more advanced than the "standard" shared memory scheme we see with Intel CPUs/iGPUs...?
I could see the argument for Apple using AMD GPUs in the first gen ASi Mac Pro, but I would think getting all the xPUs on Apple silicon should be a goal for Apple...?
Way back when Apple announced the 2018 Mac mini, with the 2019 Mac Pro not even in sight; there were distant rumblings towards Arm-based Macs...
I theorized as towards an Arm Mac Cube with a backplane offering several daughter cards; CPU, GPU, Neural Engines, video I/O, audio I/O & DSP, M.2 NVMe RAID, etc. ...
The object on stage during that 2018 keynote also pushed me towards a Cube device...
This would be quite handy if this brings support for current generation AMD GPUs - it will assist the owners of 7,1 Mac Pro and maybe even those using 5,1 if the GPUs physically fit inside their machines and the firmware can be flashed or maybe it doesn't even need that, perhaps that is limited to 6600/6800/6900.
That last bit we will see later down the track if MacOS support comes for 7900XTX and derivatives.
Last edited:
Moderator Note:
Post have been deleted, edited. Warning: keep the discussion civil or else.
Post have been deleted, edited. Warning: keep the discussion civil or else.
WHAT WE KNOW SO FAR...! ;^p
- ASi Mac Pro may use same chassis as 7,1 Mac Pro
- ASi Mac Pro may have M2 Ultra/Extreme SoCs
- ASi Mac Pro SoCs may be clocked at 4.2GHz
- ASi Mac Pro may have six PCIe slots
- ASi Mac Pro may use AMD GPUs
- ASi Mac Pro may use ASi GPUs
One interesting piece of evidence is this recently published Apple patent: https://patentscope.wipo.int/search/en/detail.jsf?docId=US392325783&_cid=P10-LFR4QS-53728-3
It describes a combination of two dies (similar to M1 Ultra) and integrates many other recent Apple patents (such as a new on-chip network, new memory folding and addressing scheme, and a new interrupt propagation protocol). Given that this patent is highly specific and combines multiple recently published technologies in one document, I believe this is likely to describe the upcoming M2 Ultra. It is common for Apple to publish patents just prior to release of new technology (this also happened with new A15 GPU and the M1 Ultra).
Note that this patent does not mention combinations of three or more dies (Extreme). I think it would be very strange for such an advanced and specific patent to not mention this if a product of this kind were preparing for production. This is why I don’t believe we will see the Extreme this year. Apple does mention for (and more) die configurations in other, more general patents. What’s interesting is that the idea of M2 Ultra only does synergies with other rumours we’ve heard. If it does feature a 4.2 GHz clock it will very favorably compete against the new Xeons and likely threadrippers, while likely outperforming a 7900 xtx.
If you want a really wild speculation though, Apple also published this rather weird looking patent: https://patentscope.wipo.int/search/en/detail.jsf?docId=US394062936&_cid=P10-LFR5E2-65898-3 Could this be a method to stick together two M2 Ultra boards and how some sort of super-fast interconnect between them? Unlikely, but one can wonder
Finally, regarding third party GPU support… as I’ve said before, I only see it in form of hardware virtualization for Linux. I just don’t see Apple sabotaging the GPU ecosystem they’ve been carefully growing in the last few years.
Obviously it's related to UltraFusion, but Apple uses to fill q bunch of patents ahead as cautionary proceed same way every technology corporation, but doesn't rules neither discard anything.
Cosas veredes pancho.
FYI support for multiple GPU vendors never removed from Metal, all iOS and Mac OS app using metal distributed an Intermediate Bytecode compiled by Metal GPU drivers into whatever binary that GPU expects doesn't matter if that GPU it's Intel AMD or Apple, that evidence Apple never ever had intentions to burn they GPU ships. if you come here with evidence Metal removes that Intermediate Binary for optimized native binaries that doesn't need on the fly compilation by Metal GPU drivers, i concede you apple has no intention to allow 3rd vendors inside a Mac, which crazy IMHO and dangerous as it corners Apple in an field where they are not even close to the 2nd most important GPU provider, and actually moreless tied with Intel ARC.
Even if you know something about metal you'll find metal 3 also allows mixing GPU vendors on the same workload even Said workload being able to be distributed on remote GPUs (past WWDC redshift run an demo on this with 4 Mac studio teetered to an master system).
If you have an idea on HPC electronics as soon you read flexible substrate you are aware it's not aimed as fabric interconnect, IMHO it's related to sub-components assembly (as extra usb4 ports, 10G Ethernet etc not on the same logic board as the main SOC).Could this be a method to stick together two M2 Ultra boards and how some sort of super-fast interconnect between them? Unlikely, but one can wonder
That would be perfect for an all-new Mac Pro Cube...!
I would rather see ASi GPUs, especially if Apple has a way to connect them into the whole SoC UMA dealio; need more GPU power & RAM, add an ASi GPU or two...!
My safe MP 8,1 (or MP 14,xxx) bets:
ASi m2, Ulta and ultra² (or whatever extreme is named), no overclock evidenced, while discreet RAM support seems safe too.
Some discreet dGPU coming along, likely AMD, Apple may offer an compute-coprocessor MPX module with m2 max as slave surrogate devices (like an new Xeon phi), but don't discard expect it to beat an AMD rx7900xtx, but maybe you may put 4 or 5 of these into the Mac Pro with same effect.
Form factor, prototypes are recycling that stunning cheese grater look, but I doubt it represents the final product, as an ASi Mac Pro whichever solution set in place would remove the room required by the Intel xeon.
System power tdp won't exceed that of '7,1 or ~1200W, i believe to be sub KW.
Arriving soon or at least before next wwdc.
ASi m2, Ulta and ultra² (or whatever extreme is named), no overclock evidenced, while discreet RAM support seems safe too.
Some discreet dGPU coming along, likely AMD, Apple may offer an compute-coprocessor MPX module with m2 max as slave surrogate devices (like an new Xeon phi), but don't discard expect it to beat an AMD rx7900xtx, but maybe you may put 4 or 5 of these into the Mac Pro with same effect.
Form factor, prototypes are recycling that stunning cheese grater look, but I doubt it represents the final product, as an ASi Mac Pro whichever solution set in place would remove the room required by the Intel xeon.
System power tdp won't exceed that of '7,1 or ~1200W, i believe to be sub KW.
Arriving soon or at least before next wwdc.