MP 7,1 Mac Pro 8.1 - WHAT IF...

[innerproduct]

Are you referring to the results here: https://www.cgdirector.com/redshift-benchmark-results/?

If so, a single RTX4080 will get you 1:47, a single RTX4090 1:16, and dual RTX4090's 47 secs. At scan.co.uk currently, RTX4080's are about £1250; RTX4090's £1750. So Apple will have their work cut out to approach the value proposition of a PC for 3D rendering performance. Using the cost of a fully-loaded 2019 Mac Pro for the comparison is soft-balling Apple somewhat.

Personally, I think I'm heading in the direction of a Studio Max, though will wait for the Mac Pro announcement in order to understand Apple's strategy for high-end ASi desktops. I'd also prefer to get an M2 Studio; apart from it generally being a good idea to skip first-gen Apple products, it will also confirm Apple's commitment to the model after a string of one-off wonders - including rumours now of the Studio itself being a holdover model until the Mac Pro's release.

Got the benches from redshift /maxon forum. Been a paying user since 2018. (But for PC)
Mostly curious what is in the cards just like you say. In the end it is a matter of cost/benefit for me. I hope for the mp2019 pricing to have been an absurd anomaly and getting back to systems with a lot better price perf ratio. Similarly to how Apple more or less have solved video editing for most of their products.

 

[avkills]

I am curious (other than cost), why PC users in general do not get "workstation" cards. You would think that anyone doing heavy 3D would be getting RTX A6000s. To me that would be a better test to the W6***X MPX cards for the 2019 Mac Pro; although even the MPX ones are built around the "consumer" GPU core. Wonder why Apple didn't want the Instinct branded AMD ones.

 

[mode11]

I am curious (other than cost), why PC users in general do not get "workstation" cards. You would think that anyone doing heavy 3D would be getting RTX A6000s.

Well, cost is the reason, really. They can be a lot more expensive. Workstation cards differ mostly in drivers, which are optimised / certified for DCC packages (and often OpenGL). They also offer features like Genlock support, ECC RAM and 10-bit colour, which depending on what you're doing, may not be important. They are often slightly down-clocked, to use less power and allow them to have slimmer, blower coolers so several can fit into a workstation. In terms of raw horsepower, though, consumer cards aren't much different; bit like the difference between an i7 and a Xeon.

To me that would be a better test to the W6***X MPX cards for the 2019 Mac Pro; although even the MPX ones are built around the "consumer" GPU core. Wonder why Apple didn't want the Instinct branded AMD ones.

Presumably because it's cheaper for them to use consumer GPUs. Unlike on the PC where there's a strict segmentation between gaming and workstation cards / drivers, on macOS GPUs are basically whatever Apple says they are.

 

[goMac]

They also offer features like Genlock support, ECC RAM and 10-bit colour, which depending on what you're doing, may not be important.

Apple has offered 10 bit color drivers for consumer cards for quite a while. But I think on Windows, with the rise of 10 bit consumer displays, 10 bit color has also started appearing on consumer cards.

 

[avkills]

Intel just released new Xeons....

AnandTech

Up to 56 cores, 8 memory channels, 4TB memory ceiling, 112 PCIe 5.0 lanes from the CPU.

Damn I wish I could sock one of those into my 2019 Mac Pro; I would probably still go for the 16 core or 20 core. Pricing does not seem as terrible as it used to be. 20 core $1989. But then again that is crazy pricing compared to AMD Threadripper.

Lower end ones up to 24 cores and 56 PCIe 5.0 lanes.

 

[ZombiePhysicist]

'Sapphire Rapids' Unveiled: Intel Xeon W-3400 Workstation CPUs Pack Up to 56 Cores

Intel's new high-performance workstation processors feature giant core counts to drive performance higher than ever before.

www.pcmag.com

 

[257Loner]

Intel just released new Xeons....

AnandTech

Up to 56 cores, 8 memory channels, 4TB memory ceiling, 112 PCIe 5.0 lanes from the CPU.

Damn I wish I could sock one of those into my 2019 Mac Pro; I would probably still go for the 16 core or 20 core. Pricing does not seem as terrible as it used to be. 20 core $1989. But then again that is crazy pricing compared to AMD Threadripper.

Lower end ones up to 24 cores and 56 PCIe 5.0 lanes.

'Sapphire Rapids' Unveiled: Intel Xeon W-3400 Workstation CPUs Pack Up to 56 Cores

Intel's new high-performance workstation processors feature giant core counts to drive performance higher than ever before.

www.pcmag.com

How on earth will the new Mac Pro compete with that?! I know Apple's SoCs are awesome for mobile computers, laptops, and small desktops, but workstations is where Intel and AMD are king. And according to those articles, Intel's workstation processors are becoming kinglier.

 

[smckenzie]

Yes, those Xeons sound pretty tasty. Even the lower tier ones are appealing. Like others have mentioned, how an earth will AS get anywhere near those.

 

[deconstruct60]

Intel just released new Xeons....

AnandTech

Up to 56 cores, 8 memory channels, 4TB memory ceiling, 112 PCIe 5.0 lanes from the CPU.

Damn I wish I could sock one of those into my 2019 Mac Pro; I would probably still go for the 16 core or 20 core. Pricing does not seem as terrible as it used to be. 20 core $1989. But then again that is crazy pricing compared to AMD Threadripper.

Crazy Pricing? Trying to compare these to the Threadripper 2000-3000 'HEDT desktop' versions? These are not those on either side; AMD or Intel.

Threadripper Pro 5000 Model specs here ( buy/shopper icons for 24 and 32 cores models )

Ryzen™ Threadripper™ Processors

www.amd.com

AMD 5965WX 24 cores $2699
AMD 5975WX 32 cores $3199

From Anadtech Article

W-7 3455 24 cores $2489 [ lower than 5965WX ]
W-7 2495X 24 cores $2189 [ lower than 5965WX ]

W-7 3465X 28 cores $2889 [ lower than 5975WX short 4 cores ]
W-9 3475X 36 cores $3739 [ higher than 5975WX up 4 cores ]
( mid point between those two price points $3,314 just $115 off the 5975WX price. )

Intel has priced these alongside AMD 5000WX Pro prices pretty closely.

The W-3400 series is going to be 'bad' pricing once drop out of the 20 cores range. You have four tiles/chiplets, each with 14 cores on them contributing to the product. The PCI-e provision ability is spread out over each tile so to get to that 112 PCI-e lane allocation have to use four tiles. Once drop into the 16 or 12 core SKUs you are only using 3 or 4 cores per die where they are 14 present. 70+ % of the cores are 'dead weight' silicon that are going to have to pay for somehow.


The W 2400 series is about as bad because once again have a extremely large monolithic die with a boatload of cores ( 34 cores ) and once get to the 50-60-70% unused stage they just aren't going to be able to price match AMD.


[
Newegg prices at the moment

AMD 5955WX 16 cores $1299 ( num of dead cores on die(s) : 0 )
W-5 3435X 16 cores $1589 ( num of dead cores on die(s) : 40 ) 'dead weight' price overhead
W-5 2465X 16 cores $1389 ( num of dead cores on dies(s) : 18 ) but is close in price (~ $100 )

Intel is likely selling those at much thinner margins than you think. it is going to be a fiscally painful 12-18 months of trying to compete with these products in the lower parts of the range and the > 60 core one. The middle is a bit less so. When Intel went to full scale production of this stuff the Data Center business profit margin shrank dramatically. This is a partially likely contributing to that.

]




I don't think Intel really wants to match AMD prices at the lower end of this range even if didn't have a 'too chunky chiplet' problem. The mainstream line up has 16-24 cores. Yeah they are 'E' cores but for non AVX-512 multithreaded code that just wants to spread the load around.... that works if don't need the I/O. Whether these are success or not is probably firmly grounded in how well the W-7 3400 and W-5 2400 products do. If they manage to compete decently (not dominate but get steady sales) , then this works as a stopgap. The W-5 3400 and W-3 2400 probably to keep some targeted fringe market customers happy with at acceptable losses. Sales of the W-9 3400 and W-7 2400 will help the losses be a bit more tolerable ( they can't save the product line, but could help the fiscal numbers be less dire. )

The sales pitch on the 3400/2400 is either where hit a sweet spot computationally with AVX-512 , tap into one of the accelerators , and/or have a bigger PCI-e I/O bandwidth problem than a core count problem. To bring it back to a Mac Pro context. Apple skipped AVX-512 in Rosetta ( pretty indicative that they don't really think it is strategic or tactically relevant), Intel proprietary accelerators (never did anything with the RAID features in the I/O controller the decade used Intel parts. So why would they care now? Nope ), More than 64 lanes the Mac Pro really doesn't urgently need ( and being under 64 lanes is even less urgently needed. )

And macOS still doesn't do more than 64 CPU threads. So the W-9 3400 series doesn't do much there either.

From a technical perspective , I don't think Apple would want to touch these with 10-foot-pole. When AMD releases the Threadripper 7000WX late in 2023 these are going to get toasted in about every way possible in terms of Pref/$. That is a lot of work to upgrade only to get smoked in less than a year. Nevermind they'd have to completely backpedal from their Apple Silicon is the future after 2022 stance. In the 20-28 CPU core range they have a better solution.

Pricing wise it is going to be much bigger problem pushing the $2-3K , '> 1TB RAM' tax on people when it is completely missing from the AMD and Intel offerings now which are substantially more performant also.



Intel has an odd approach to their accelerator 'value add' .

The three 24 - 28 products above

Intel® Xeon® w7-3465X Processor (75M Cache, 2.50 GHz) Product Specifications

Intel® Xeon® w7-3465X Processor (75M Cache, 2.50 GHz) quick reference guide including specifications, features, pricing, compatibility, design documentation, ordering codes, spec codes and more.

ark.intel.com

Intel® Xeon® w7-3455 Processor (67.5M Cache, 2.50 GHz) Product Specifications

Intel® Xeon® w7-3455 Processor (67.5M Cache, 2.50 GHz) quick reference guide including specifications, features, pricing, compatibility, design documentation, ordering codes, spec codes and more.

ark.intel.com

https://ark.intel.com/content/www/u...eon-w72495x-processor-45m-cache-2-50-ghz.html

All only have the Data Streaming accelerator turned on. (again more 'dead cores' riding around on the die(s) ).

Lower end ones up to 24 cores and 56 PCIe 5.0 lanes.

Which basically blocks it from being good solution for a Mac Pro update using the 2019's baseline model of 64 lanes from the CPU package being available. it is a better thermal cooling fit though. ( 225W vs 300W both of which Intel is going to blow way over if push these hard. )



P.S. before in the Xeon high end Server processor space Intel would do a XCC (extreme core count) , MCC (medium core count) , and LCC ( 'low' core count) die and place all three into a single package. Here there are two tile dies (mostly overlapping but asymmetric ) and one humongous monolithic die oddly being named "MCC" when it is even more monstrous die area wise than the old XCC dies. Still three dies all of which are widely more complex that before ( accelerators (which mainly just work in server space) , CXL 2.0 , etc. )

Intel globbed way too much stuff into one 'catch up in single tremendous leap' iteration. Pulling up way too much complexity usually doesn't work. And exactly why Apple shouldn't be trying to snuff out every single possible usage of the old Intel Pro in one single leap with Apple Silicon. It is a long term race. Do part of the range. Then do the next part.

 

[deconstruct60]

How on earth will the new Mac Pro compete with that?! I know Apple's SoCs are awesome for mobile computers, laptops, and small desktops, but workstations is where Intel and AMD are king.

Apple Silicon had an 'AMX' before Intel had 'AMX'. That is one way they are going to complete.
The other part is that just aren't going to follow them into the > 64 PCI-e lane space or > 64 core space. 20 'about as fast' cores at lower prices would help. Apple doesn't have to drop clock speeds when the crank up the active core count. These produccts do. So in the 10-25 CPU core count range they compete decently.

NPU throughput ... Apple isn't that far 'behind'.

And if Apple had a 40 CPU core product it would compete over an even wider range on CPU/NPU workloads.

And according to those articles, Intel's workstation processors are becoming kinglier.

Kingly ? Errr. not really. A fading king perhaps. AMD is only going to grab more market share this year. This stuff isn't going to completely stop the slide. It may slow the slide down somewhat as AMD Threadripper WX is creeping up on being a year old at this point and still has the very top end performance with their 64 core offering. The middle is more competitive. (and Intel is loosing its shirt in profits at the lower end of this range. See article above).

 

[avkills]

I sure hope Apple has a hush-hush surprise for us soon.

My gut feeling is that the AS Mac Pro is going to have its own silicon; it seems the only way they could have PCIe expansion and adequate number of Thunderbolt & USB ports.

Perhaps there will be a T2 type system controller that handles the mundane ports and CPU/GPU modules that slot in (that add additional Thunderbolt/Displays and RAM; and CPU/GPU grunt.)

 

[deconstruct60]

I am curious (other than cost), why PC users in general do not get "workstation" cards. You would think that anyone doing heavy 3D would be getting RTX A6000s. To me that would be a better test to the W6***X MPX cards for the 2019 Mac Pro; although even the MPX ones are built around the "consumer" GPU core. Wonder why Apple didn't want the Instinct branded AMD ones.

Apple did pragmatically use the "Instinct" like "Pro" cards with the Pro Vega II MPX. When AMD was borrowing money to keep the lights on their consumer/datacenter Pro GPU lines were pragmatically the exact same thing because they didn't have money or resources to do anything else different.

AMD and Nvidia data center HPC/AI/ML compute cards though typically dump the display controller to save space that is reallocated to things like FP64. ( Apple is going the other direction where doesn't see usefulness of even implementing FP64 as a native hardware supported datatype. ).

As CDNA and RDNA have split more at 2nd generation stage there are more difference. One of which is video out.

CDNA3 will somewhat split even more by weaving Zen x86 cores into the Instinct card in a more unified memory fashion. Although is suspect there will be a PCI-e card variant that either downgrades the x86 core count or drops back to a more traditional set up just to hit a 300-400W thermal envelope that would better fit in a workstation. All the variants can't be super expensive and regular PCI-e slot thermal nightmares in a generic tower 'box'.

The A100 would be more lined up with what the Instinct brand is providing on the AMD side.

Instinct may line up with a future Mac Pro if Apple is a bit more tolerant of asymmetric compute cards. ( MP 2013 either got two D300 , two D500 , or two D700 . 6800 Duo but not connected path to 6800-6900 combo . The Ultra is an exact twin Max only . ). Instinct card do treat FP64 as a native hardware datatype , do ECC VRAM , have a different set of matrix engine options than Apple ---- all things Apple's GPU doesn't do.

If Instinct was more a "supercomputer node on a card" with CDNA3 even more different augment of what Apple doesn't do.


The A6000 is more a "Pro" card with some graphics output still a major focus. The difference is more so drivers. Mainstream GPU drivers go through tons gyrations chasing quirks in an endless list of 'new games'. So they don't really seek stablity as much as "got a xx% uplift in trendy game XYZ this quarter" and shift to the newest 'shiny' quarter after quarter following that.

The "pro" drivers are usually optimized across only a reasonably sized, and relatively much smaller, subset of applications. A longer , usually far more steady evolutionary path that doesn't chase the latest shiny app. From time to time "pro" drivers may have a narrow regression , but typically no where near as many as the mainstream drivers do.

One upside of macOS graphics stack not having a 'latest shiny , churning gaming market to chase after' was that the macOS graphics drivers tended to follow a path the "Pro" ones do on the Windows side. Likewise for a less choatic GPU instances also. The subset of GPU choices that Apple embedded into their Macs made 'which' optimizations go 'where' a much more tractable exercise to do in a far more focused fashion. (as opposed to chasing everything for everybody).

What was not seeing on Windows "mainstream" vs "Pro" split was a paradigm split. Where had one radically different memory allocatoin and latencies on on side and the other market following something different. The application's approach to using a A6000 and mainstream variant was the same. How much data could get cached in VRAM could be different, but the primary optimization approach was the same.

 

[deconstruct60]

I sure hope Apple has a hush-hush surprise for us soon.

My gut feeling is that the AS Mac Pro is going to have its own silicon; it seems the only way they could have PCIe expansion and adequate number of Thunderbolt & USB ports.

Different from all of the other laptop dies. Probably. The laptop dies are all monolithic. The Mac Pro (and Studio) probably are not. Because they are plugged in all the time they can give slightly on Perf/Watt and take the hit of having 1-4 UltraFusion connectors in the package for a relatively very small Pref/Watt retreat.

Perhaps there will be a T2 type system controller that handles the mundane ports and CPU/GPU modules that slot in (that add additional Thunderbolt/Displays and RAM; and CPU/GPU grunt.)

Neither the Intel W-3400/2400 nor AMD Threadrippers need a 'slot' to provision PCI-e lanes out of the respective packages at all.

Perhaps The 'T2' part on a chiplet that only need to add one per package, but there is no need to 'slot' it at all.


Take a Max Die and decompose that into two or three chiplets. For example

Top edge I/O -- SSD , Secure Enclave , USB , x1 PCI-e v4 lanes for basic Network/etc , HDMI/eDP out , four TBv4 controllers , one UltraFusion edge to connect back to mesh.

'the rest' -- memory controllers and System Cache blocks, CPU cores , NPU core , GPU , DisplayControllers , A/V de/encode, two UltraFusion edges; one on top and bottom

'bottom' edge I/O - two x16 PCI-e v4 , 1 or 2 x4 PCI-e v4 bundles and maybe two TBv4 controllers (or an extra SATA/USB controller).


With those three building blocks could make the following products.

1. Mn Max with 6+ TB ports. [ top edge i/o + 'the rest' + flipped top edge i/o connected to bottom 'the rest' connector]

12+ CPU cores 38+ GPU cores etc.

Could put this in a Mac Studio where don't have to 'downgrade' the front two USB ports to non Thunderbolt. ( slightly wasteful in that have extra SSD controller , extra two Thunderbolt , extra HDMI/eDP out , but relatively small 'dead space' )

2. Mn Ultra S with 6+ TB ports . [ top edge i/o + 'the rest' + 'the rest' + flipped top edge i/o connected to bottom 2nd 'the rest' ]

24+ CPU cores , 38+ GPU cores , etc.

Again could put this in a Mac Studio just like the M1 Ultra went into one.

3. Mn Ultra P 4-6 TB ports [ top edge i/o + 'the rest ' + flipped 'the rest ' + bottom edge I/O ]

24+ CPU core , 76+ GPU cores , 32+ PCI-e v4 lanes , etc.

If this have 6 TB ports could merge this with Ultra S, but if need to drop unused TB ports to get PCI-e lanes, that is a better option. That would gap the Studio and Mac Pro


4. Mn 3-deep P [ top end i/o + 'the rest' + flipped 'the rest' + 'the rest' + bottom edge I/O ]

36+ CPU cores , 114+ GPU cores , 32+ PCI-e v4 lanes , etc.

Probably wouldn't scale as well as the paring with the first two large GPU clusters closely adjacent. And last two couplings would hat two CPU groups adjacent but the third section also more distant. The on package workload direction first light up those two adjacent groupings first. When the workload spills bigger than an Ultra then fire up the more distance addition units of that type. Also likely easier if a more distance GPU cluster is driving 3rd or 4th monitor and main cluster just tasked when have single primary display.

30 CPU cores would hang much better the middle of W-3400 or 5000WX line ups. 100+ cores is likely enough to top a W6800X on more than a few workloads. And the RAM memory cap is incrementally higher (along with more bandwidth available)


5. Mn 1-deep P [ top end i/o + 'the rest' + bottom edge I/O ]

Probably not interesting to high end users, but an incremetnally more affordable 'hobby box' to toss PCI-e cards into would work. [ If Mac Pro only had 4 TB sockets wouln't necessarily need bottom edge I/O to provide the final 2. ]

[or if tighted up the thermals a bit could put this on a add-on PCI-e card for a "Mac Card" . SoC , small drive , Ethernet and display ports. If under 75W doesn't even need power connector. ]








It would help for 'the rest' modules to be TSMC N3 so that the totally distance 'top' to 'bottom' across the middle was substantively shorter. The smaller the gap the more likely it will present as 'unified and uniform' collection of cores.


The collection of silicon is different (from laptops) and yet not completely different. Some amount of reusability is likely plays a major role. It is the only pragmatic way of keeping the costs affordable.

 

[TrevorR90]

The newly announced xeons are the only CPU I would like in the new Mac Pro.

Wishful thinking on my part but apple silicon in its form right now, doesn't make sense for the Mac Pro IMO.

 

[avkills]

If they do manage PCIe expansion on the Apple Silicon Mac Pro, they should go straight to v5 and skip v4 IMO.

 

[innerproduct]

I really hope that Apple just been good at keeping their secrets and that @maikerukun will be right in that there is a super duper chip coming up for us all to be amazed by. All in all, the XEON has few things going for it except the PCI lanes. There are already some leaked benches for the 24 core version and it doesn't impress. It might be faster (than a ASi core based machine) in some specific task but over all, it is not an exciting product.
I return to the conclusion that Apple have known for a long time how the Intel roadmap looks and also what AMD would be able to produce both when it comes to GPUs as well as CPUs.
The only sane conclusion is that Apple will release a bad ass system based on ASi. Maybe they annouce it within a few weeks and then we can all wait for delivery a few months...
Absolute dream system: something beautiful in the vein of the NeXT cube, the Apple Cube and the thrashcan but with user replaceable "Pro Blades" fanning out from an ultra fusion core. CPU cores 12-48. GPU cores: 38-304. Some kind of expandable RAM/cache whatever.
I want to have the feeling of approaching an alien artefact when I see it. Both for its beauty but also for its power.

 

[Kimmo]

Interesting.

Future Mac Pro may use Apple Silicon & PCI-E GPUs in parallel | AppleInsider

Despite Apple Silicon currently working solely with its own on-board GPU cores, Apple is researching how to support more options, like PCI-E GPUs, all working in tandem.

appleinsider.com

 

[Boil]

Interesting.

Future Mac Pro may use Apple Silicon & PCI-E GPUs in parallel | AppleInsider

Despite Apple Silicon currently working solely with its own on-board GPU cores, Apple is researching how to support more options, like PCI-E GPUs, all working in tandem.

appleinsider.com

To my layman mind, this seems to be an indicator that we may see ASi GPGPUs as an option for more compute/render horsepower in the forthcoming ASi Mac Pro, and we could also see Apple offering these ASi GPGPUs in external housings, meaning e(GP)GPUs for the slotless ASi Macs...

But I would think initial launch would keep the ASi GPGPUs to the ASi Mac Pro, supply constraints and all that...?

 

[edanuff]

To my layman mind, this seems to be an indicator that we may see ASi GPGPUs as an option for more compute/render horsepower in the forthcoming ASi Mac Pro, and we could also see Apple offering these ASi GPGPUs in external housings, meaning e(GP)GPUs for the slotless ASi Macs...

But I would think initial launch would keep the ASi GPGPUs to the ASi Mac Pro, supply constraints and all that...?

Yeah, this seems like the missing piece of the pie at last. Apple GPU cards and even possibly external expansion as well as a GPU in the new monitors.

 

[deconstruct60]

Anandtech ...
....
Lower end ones up to 24 cores and 56 PCIe 5.0 lanes.

Actually, looping back over the spec sheets and some other detailed write ups. The W 2400 series has 64 PCI-e v5 lanes. Earlier rumors had put it at 56 so I glanced over that number and didn't notice.

Kind of odd is that there are 5 x16 PCI-e controllers according to some "MCC" die floorplans / die shots. So have a 'spare' ( and probably a better match to the motherboards anyway given the chip is so large; most boards are not going to have more than low single digit number of slots with direct connection to the CPU package. ) . [ I suspect there was going to be a smaller die at some point which was an actual classic 'MCC'.

New Workstation Kingpins Intel Xeon W-3400 Xeon W-2400 and W790 Launch

We delve into the new Intel Xeon W-3400 and Xeon W-2400 series bringing PCIe Gen5 lanes, DDR5, and more cores to the workstation market

www.servethehome.com

Take out the middle column with the UPI line at the top. Take out the left mode column with the rest of UPI links don't need at all (or shift the top one over and take out the '5' PCI-e controller). Shift the memory controllers into the core grid positions in the column to the right. Do the same thing on rightmost column ( dump half of the accelerators not going to turn on anyway and shift memory controllers left to remove some cores.. You go from a 34 cores down to 17 . Still kind of loppy to do a 6 and 8 core part out of that, but far less loopy than out of a 34 core die. I think Intel didn't have time for that with rash of bugs that sprung up around these designs and just pushed the biggest dies out the out the door and will just have to 'made do' with that.
]
)


It is less of a package mismatch with the MP 2019 board. Fewer memory channels ( don't think Apple would blink at dropping some DIMM slots for a stop gap Intel Mac Pro needed to do more inexpensively. In fact, could probably drop down to just 4 DIMM slots to simplify the new board and validations. ). Prices better. Thermal better. (relative to MP 2019). If Intel had launched that in late 2020 or even 1H 2021 Apple might have taken that (if that Studio and Mac Pro AS models were going to slide into Q2 2022 and Q? 2023. ). Up against the Threadripper W2000 in 2021 this would have been very competitive. ( instead all Intel had was a demonstratively worse W-3200 product. )



Of course, if Apple did a W-2400 Mac Pro instead of a 3400 based (or Threadripper) one there would still be 'boo birds' saying it was horrible backslide and so not worthy of 'Mac Pro' label. 'The core count had to go up'. The Max Memory had to go up or else colossal 'fail'. etc. etc.

 

[mcnallym]

Interesting.

Future Mac Pro may use Apple Silicon & PCI-E GPUs in parallel | AppleInsider

Despite Apple Silicon currently working solely with its own on-board GPU cores, Apple is researching how to support more options, like PCI-E GPUs, all working in tandem.

appleinsider.com

Would kind of make sense. Keep the GPU cores in the SoC foe actual Screen Drawing etc and then have additional compute GPGPU for tasks like rendering etc

Perhaps hiding behind something like how use VideoToolBox to allow use of Media Engine on AS or QuickSync on Intel Macs so that the Mac can use what is available

 

[Boil]

Would kind of make sense. Keep the GPU cores in the SoC foe actual Screen Drawing etc and then have additional compute GPGPU for tasks like rendering etc

Exactly what I have been saying for many many months now...

The trick is for Apple to have variants; cards with two, four, or eight "GPU-specific" SoCs (and appropriate amounts of RAM for each); and for the pricing (especially considering "eGPU" variants and the cost of the external housing/PSU) to be reasonable enough for folks to consider these for use with their slotless ASi Macs...

Economy of scale and all that...

 

[Oculus Mentis]

How on earth will the new Mac Pro compete with that?! I know Apple's SoCs are awesome for mobile computers, laptops, and small desktops, but workstations is where Intel and AMD are king. And according to those articles, Intel's workstation processors are becoming kinglier.

In a parallel universe Apple would make peace with nvidia and release a Mac Pro based on their new Grace Hopper cpu+gpu architecture, which should be binary compatible with AS since are both based on arm…

 

[freqrider]

What if the new macPro just used off the shelf AMD EPYC or similar workstation grade chips, forgoing SOC altogether, and kept upgradability? That would be my best guess.

 

[Hastings101]

What if the new macPro just used off the shelf AMD EPYC or similar workstation grade chips, forgoing SOC altogether, and kept upgradability? That would be my best guess.

I think that would be quite the PR hit for Apple, but so would just using an "M2 Ultra" as the rumors suggest is going to happen.

What if Apple just delays the Mac Pro for years and years until they can finally compete with their own silicon at the pro Workstation and Server level