The 2013 Mac Pro has a 450 W power supply, just a little bit less than the 2017 iMac Pro. Most of the extra 50 W in the iMac Pro went to powering its display, and the two have about the same power envelope for the computing bits.
Apple lists the thermals for each, it's probably easier to cool the iMac Pro than the cylinder but they clearly wanted to hit a much higher power level in the Mac Pro:
This was with Intel/AMD parts though, which are around 5x less power efficient than Apple Silicon. While they could target the same power level, to do that would need bundling around 10 x M1 Max chips, which would be over 100 TFLOPs of computing power. That seems excessive for what the workstation market needs in a single machine. I expect they will instead target the same performance level at a lower power.
On 3nm, they can get 70% more transistors than 5nm so potentially over 60TFLOPs in under 300W. This is the performance level the next-gen 5nm GPUs are reported to have:
Someone could have quadruple those but even today it's rare that people buy more than a single 3090. There a video here where they put 4 x 3090 in an external box:
That's the kind of PCIe box Apple could make for people who need it for non-GPU peripherals. A Mac Pro chip on 3nm would perform roughly like the 2x 3090s in that video. There's also nothing wrong with buying PCs for this kind of task, Apple doesn't have to cater to every use case. Someone can buy a Mac Pro for the creative part and offload the render to a farm of PCs with multiple GPUs.
If it is a diminishing market, they wouldn't be selling a 2019 Mac Pro, right? It seems more that they didn't understand the workstation market, outside of a narrow video editor one. They didn't anticipate machine learning, I think they gave up on CAD and 3D in that 2012 to 2015 period. Computational codes (anything involving large sets of differential equations) didn't interest them that much? There is a market out there, where computational requirements are basically infinite, so my expectation is the workstation market is here to stay for the foreseeable future.
The 2019 Mac Pro was requested by some of their high profile Mac users:
Tim Cook or one of the other execs said that the Mac Pro segment doesn't move the needle for them, it's just a space they decided they wanted to cater to. It works as a halo product. Some of the people who requested this model have businesses that sell PCIe cards so they have a financial incentive for these models to still be sold:
Mac Pro buyers only number in the tens of thousands per year but if even 2000 of those buy a $1000 PCIe card, that can make an entire business that sells those cards viable and could cause some businesses to go bankrupt without that support.
The market with infinite computation requirements doesn't have to do it all on a single machine. A single workstation only has to be good enough for real-time tasks like 8K video and offer good performance-per-dollar. Offline computation can be done with multiple machines.
It's great that the 2021 MBP machines have Mac Pro performance, but that doesn't mean it ends there. It only means people want 2x to 10x the performance of the 2021 MBP in a desktop box. My perspective is that there are a lot of use cases that have continually increasing compute needs, and therefore, those customers will continually want the most compute performance that can be placed in a desktop box with a 1000 W of power.
HP is the biggest manufacturer in the workstation market, their 2020 annual report is here:
70% notebook, 25% desktop, 5% workstation. For Apple it's closer to 80% notebook, 19% desktop, < 1% workstation.
Globally the unit volume of workstations is reported in different places around 1 million units per quarter of which Dell and HP are the large majority and split evenly, something like 40% each. Assuming HP is selling 1.6m workstations per year, this makes their ASP $1125 per workstation. That means models like these that start around $1200 that have things like Core i5 CPU, 8GB RAM:
Those models are recommended by HP for things like architecture planning. The high-end workstation market ($3k+) is a fraction of the overall workstation market and it will get smaller over time the more that people migrate to notebooks, which can now do the jobs workstations were used for.
I'd say the premium segment ($3k+) would likely be no more than 25% of the overall volume so Apple would be competing in a market of around 1m units per year. They won't be able to beat HP/Dell because they are used in the server space but they could possibly get 20% of the premium market at around 200,000 units per year (still < 1% of Mac unit volume). At an ASP of $5k, this is $1b.
Macs last year made $35b, iPhone $191b, iPad $31b, wearables $38b, services $68b. They make way more from even AirPods and the Apple Watch than a best-case for workstations.
If the workstation market decides they prefer external GPUs with add-in boards, Apple would be better just leaving the Intel Mac Pro as is and bump the spec every now and then and put Apple Silicon in the iMac Pro. In about 5 years, there will be no more upgrades to go to and no more demand for premium desktops.
The 2013 Mac Pro has a 450 W power supply, just a little bit less than the 2017 iMac Pro. Most of the extra 50 W in the iMac Pro went to powering its display, and the two have about the same power envelope for the computing bits.
Apple lists the thermals for each, it's probably easier to cool the iMac Pro than the cylinder but they clearly wanted to hit a much higher power level in the Mac Pro
Yes. This is probably why Federighi said thermal corner - it was interesting that Federighi got the "thermal corner" money quote and not Ternus or Schiller who would have been the people most responsible for the decisions. However, I don't think the iMac Pro form factor makes it any easier to cool 500 Watts of components than the 2013 Mac Pro form factor. The total power consumption of the two machines is basically the same, and the cylinder is totally capable of cooling 500 W of components. For unknown reasons, Apple thought the 2017 iMac Pro was a good way to address the high end of the market, and I think cooling had very little play there. They thought an AIO was going to be good enough.
Then, yes, they got feedback that that the 2017 iMac Pro wasn't going to be enough and PCIe slots are an important feature for some of their big customers, however else they arrived at the decision, and the 2019 Mac Pro was a go and they said it will be coming.
This was with Intel/AMD parts though, which are around 5x less power efficient than Apple Silicon. While they could target the same power level, to do that would need bundling around 10 x M1 Max chips, which would be over 100 TFLOPs of computing power. That seems excessive for what the workstation market needs in a single machine. I expect they will instead target the same performance level at a lower power.
If Apple can put 100 TFLOPs in a Mac Pro, at less than 1000 W, they should. People will buy it for their desks. People buy lots of them and rack them. Don't understand the hesitancy here. If Intel, Nvidia or AMD could, they would. You design in performance up to all practical limits for the given form factor. If the half sized Apple Silicon Mac Pro offers 40 TFLOPs at 250 W, well Apple should design in a high bandwidth network for them so people can buy 2, 3, 4 of them on their desks or have lots of them racked. They have to get on with getting HPC codes converted to Metal.
I'm not a proponent of "which came first, the chicken or the egg" type situations btw. The hardware always comes first, and it's always up to the OEM to sell a lot of units with the right customers for developers to target. The developers are always fast followers. So ultimately, Apple has to offer a reasons above and beyond just performance. So, Metal, Metal, Metal conversions of HPC code, which they will likely have to pay for. If they offer a lot of CPU cores, they have to get HPC code optimized for ARM.
Mac Pro buyers only number in the tens of thousands per year but if even 2000 of those buy a $1000 PCIe card, that can make an entire business that sells those cards viable and could cause some businesses to go bankrupt without that support.
The market with infinite computation requirements doesn't have to do it all on a single machine. A single workstation only has to be good enough for real-time tasks like 8K video and offer good performance-per-dollar. Offline computation can be done with multiple machines. ... I'd say the premium segment ($3k+) would likely be no more than 25% of the overall volume so Apple would be competing in a market of around 1m units per year. They won't be able to beat HP/Dell because they are used in the server space but they could possibly get 20% of the premium market at around 200,000 units per year (still < 1% of Mac unit volume). At an ASP of $5k, this is $1b.
Macs last year made $35b, iPhone $191b, iPad $31b, wearables $38b, services $68b. They make way more from even AirPods and the Apple Watch than a best-case for workstations.
If the workstation market decides they prefer external GPUs with add-in boards, Apple would be better just leaving the Intel Mac Pro as is and bump the spec every now and then and put Apple Silicon in the iMac Pro. In about 5 years, there will be no more upgrades to go to and no more demand for premium desktops.
I wouldn't make a bet of 5 years. It's the usual push-pull of client vs server computing. Maybe cloud computing will be so cheap, that it will drive high performance workstations down to "not worth Apple's time", but I wouldn't make that bet for 5 years. Wouldn't make it all. There will be applications gated by the network, whether by bandwidth, by security, network stability, or whatnot. So, I think there will always be something that will drive sales of workstations on desks. What was done on a compute cluster yesterday, is done on a workstation today. What is done on a computer cluster today, will be done on a workstation tomorrow. The computer cluster just gets bigger, handling ever bigger problems.
Then, I would say the notion of the "market is too small to be worth Apple's time" as short sighted. Much like the decision to have a 2019 Mac Pro, it wasn't directly about units sold. It's about platform or ecosystem stability. Every niche that Apple doesn't play in is another reason for a loss of a sale of product in their ecosystem. Eg, no workstation sale could mean no MBP sale to that customer. The more niches that Apple doesn't play in, the more it limits how many products they can sell. That Mac Pro enables sales of the MBP because it's keeping a lot of software on the macOS platform for MBP to run.
A gaming PC is a particularly easy and acute example. You do not get a Mac to play games. You get a PC. Well, that means that person may get a Windows laptop instead, or an Android phone instead. There are a lot of apps like this. Not having Creo hurts (the Mac workstation troubles probably directly plays into this). The Matlab implementation is limited. Not having something competitive to CUDA (hopefully yet) and Nvidia's class leading performance hurts. Microsoft doesn't even have all of the Office apps on Mac. Not having MS Project hurts sales of Macs. Who the heck knows what Windows only software that hospitals run? Every niche that Apple doesn't have a foothold limits how many units they can sell.
If Apple can put 100 TFLOPs in a Mac Pro, at less than 1000 W, they should. People will buy it for their desks. People buy lots of them and rack them. Don't understand the hesitancy here. If Intel, Nvidia or AMD could, they would. You design in performance up to all practical limits for the given form factor. If the half sized Apple Silicon Mac Pro offers 40 TFLOPs at 250 W, well Apple should design in a high bandwidth network for them so people can buy 2, 3, 4 of them on their desks or have lots of them racked. They have to get on with getting HPC codes converted to Metal.
Then, I would say the notion of the "market is too small to be worth Apple's time" as short sighted. Much like the decision to have a 2019 Mac Pro, it wasn't directly about units sold. It's about platform or ecosystem stability. Every niche that Apple doesn't play in is another reason for a loss of a sale of product in their ecosystem. Eg, no workstation sale could mean no MBP sale to that customer. The more niches that Apple doesn't play in, the more it limits how many products they can sell. That Mac Pro enables sales of the MBP because it's keeping a lot of software on the macOS platform for MBP to run.
That's one of the issues that comes up with new hardware models is how to make it more appealing than the last one. This affects every product they make and it gets harder each time. Even with popular products like the iPhone, they will run out of enough things to make a compelling upgrade.
This affected the 2013 Mac Pro too in that it didn't offer much more than what people already had and the higher volume users at the lower performance levels migrate down to cheaper models. Apple's Mac growth in the last few years has been driven by notebooks. If someone has a specced out 2019 model that already offers up to 64GB video memory, 1.5TB RAM, up to 56TFLOPs GPU, native Windows support for specialist software like CAD, it's hard to make a compelling upgrade in a new Apple Silicon model.
That's why it makes sense to use 3nm. This will give them up to 70% more transistors. This isn't good timing-wise though as 3nm won't be ready next year.
If M2 will be on 4nm, they can do the iMac Pro early 2022 with M1 Pro up to M1 Max Duo. Later in the year they do M2, M2 Pro, M2 Max for the MacBooks and M2 Max Duo/Quad for the Mac Pro and the iMac Pro can get a refresh. Then it's 3nm for M3 late 2023.
For compute, Nvidia has done a lot software-wise to boost adoption like with raytracing support and other AI libraries to accelerate real-world workflows. Like you say, it would help boost Metal adoption and software performance with more Metal libraries to be able to use all that compute power:
While I agree that Apple should continue to cater to high-end users so that software support is there, that doesn't have to mean a form factor like the 2019 model. There were pro users that bought iMac Pros but the audience for these high-end models in every form factor is small so every model will seem like a commercial flop. The main thing is they cater to real world workflows and with their work on Apple TV+, they have direct access to higher end visual effects and what the needs are.
I bought a 2019 Mac Pro - dual x5700, 64gb, afterburner, and 8 core (with intents to upgrade it to 24 core later). I also bought a XDR display. I knew the event was coming and bought within 12 days of the event. I honestly didn’t anticipate the new MBP being able to keep pace on 8k editing in Davinci Resolve and rendering in Cinema4d. I took it back and kept the XDR. The machine was a powerhouse, was sexy and honestly the coolest Apple product I owned in 25 years. But also comparing a $13k machine to a $3.5k machine that does about 98% of the performance, I couldn’t justify the price difference for my case. I can’t wait for the Apple Silicon Mac Pro.
This.
The Intel-based Mac Pro is essentially dead imho. I don't see how the math could be made to work to justify a Mac Pro knowing how AS is performing. It's not necessarily a function of evolution in technology per-se, but Intel's inability to step up and perform. It's embarrassing really how bad Intel looks right now. Intel is the Titanic that is continuing on without knowing they crashed into the ASi Iceberg.
Even though I'm years away from buying a new desktop Mac, after seeing what the new MBP's are doing on the performance bench, I'm super-excited to see what Apple does with the iMac and Mac Pro using ASi. With larger footprints for better cooling and larger chips, I'm expecting them to be serious Intel-killers.
It's not just that the Intel Mac Pro that's dead. The entire X86 platform is about to enter assisted living. The platform is decades beyond its expiration date and that company has done nothing, really, to keep their platform modern. And AMD Ryzen also. Soon, all the PC chip makers will be switching over to ARM, and will pretend that they're the innovators. In five or seven years, you'll be hard pressed to find an X86 PC at Best Buy.
As a photographer by hobby only now, and retired from video production, I am loving where Apple hardware is at. After the 2013 Mac Pro became obsolete I moved on to the iMac Pro. I resisted the new Mac Pro as overkill for my needs.
I have an M1 Mac mini so as to test and learn about the new tech, but work on my i9 27” iMac 64 GB. The new M1 MBPs are nice but the next logical machine for me will be the 2022 large format iMac unless there is a new Mac Pro, but even if there is I may refrain. If the 2022 iMac has a 6K screen and at least 27” or more and sports a quad M1 Max I would be very, very happy with that, for sure.
The 2013 Mac Pro has a 450 W power supply, just a little bit less than the 2017 iMac Pro. Most of the extra 50 W in the iMac Pro went to powering its display, and the two have about the same power envelope for the computing bits.
Apple lists the thermals for each, it's probably easier to cool the iMac Pro than the cylinder but they clearly wanted to hit a much higher power level in the Mac Pro
Yes. This is probably why Federighi said thermal corner - it was interesting that Federighi got the "thermal corner" money quote and not Ternus or Schiller who would have been the people most responsible for the decisions. However, I don't think the iMac Pro form factor makes it any easier to cool 500 Watts of components than the 2013 Mac Pro form factor. The total power consumption of the two machines is basically the same, and the cylinder is totally capable of cooling 500 W of components. For unknown reasons, Apple thought the 2017 iMac Pro was a good way to address the high end of the market, and I think cooling had very little play there. They thought an AIO was going to be good enough.
Then, yes, they got feedback that that the 2017 iMac Pro wasn't going to be enough and PCIe slots are an important feature for some of their big customers, however else they arrived at the decision, and the 2019 Mac Pro was a go and they said it will be coming.
This was with Intel/AMD parts though, which are around 5x less power efficient than Apple Silicon. While they could target the same power level, to do that would need bundling around 10 x M1 Max chips, which would be over 100 TFLOPs of computing power. That seems excessive for what the workstation market needs in a single machine. I expect they will instead target the same performance level at a lower power.
If Apple can put 100 TFLOPs in a Mac Pro, at less than 1000 W, they should. People will buy it for their desks. People buy lots of them and rack them. Don't understand the hesitancy here. If Intel, Nvidia or AMD could, they would. You design in performance up to all practical limits for the given form factor. If the half sized Apple Silicon Mac Pro offers 40 TFLOPs at 250 W, well Apple should design in a high bandwidth network for them so people can buy 2, 3, 4 of them on their desks or have lots of them racked. They have to get on with getting HPC codes converted to Metal.
I'm not a proponent of "which came first, the chicken or the egg" type situations btw. The hardware always comes first, and it's always up to the OEM to sell a lot of units with the right customers for developers to target. The developers are always fast followers. So ultimately, Apple has to offer a reasons above and beyond just performance. So, Metal, Metal, Metal conversions of HPC code, which they will likely have to pay for. If they offer a lot of CPU cores, they have to get HPC code optimized for ARM.
Mac Pro buyers only number in the tens of thousands per year but if even 2000 of those buy a $1000 PCIe card, that can make an entire business that sells those cards viable and could cause some businesses to go bankrupt without that support.
The market with infinite computation requirements doesn't have to do it all on a single machine. A single workstation only has to be good enough for real-time tasks like 8K video and offer good performance-per-dollar. Offline computation can be done with multiple machines. ... I'd say the premium segment ($3k+) would likely be no more than 25% of the overall volume so Apple would be competing in a market of around 1m units per year. They won't be able to beat HP/Dell because they are used in the server space but they could possibly get 20% of the premium market at around 200,000 units per year (still < 1% of Mac unit volume). At an ASP of $5k, this is $1b.
Macs last year made $35b, iPhone $191b, iPad $31b, wearables $38b, services $68b. They make way more from even AirPods and the Apple Watch than a best-case for workstations.
If the workstation market decides they prefer external GPUs with add-in boards, Apple would be better just leaving the Intel Mac Pro as is and bump the spec every now and then and put Apple Silicon in the iMac Pro. In about 5 years, there will be no more upgrades to go to and no more demand for premium desktops.
I wouldn't make a bet of 5 years. It's the usual push-pull of client vs server computing. Maybe cloud computing will be so cheap, that it will drive high performance workstations down to "not worth Apple's time", but I wouldn't make that bet for 5 years. Wouldn't make it all. There will be applications gated by the network, whether by bandwidth, by security, network stability, or whatnot. So, I think there will always be something that will drive sales of workstations on desks. What was done on a compute cluster yesterday, is done on a workstation today. What is done on a computer cluster today, will be done on a workstation tomorrow. The computer cluster just gets bigger, handling ever bigger problems.
Then, I would say the notion of the "market is too small to be worth Apple's time" as short sighted. Much like the decision to have a 2019 Mac Pro, it wasn't directly about units sold. It's about platform or ecosystem stability. Every niche that Apple doesn't play in is another reason for a loss of a sale of product in their ecosystem. Eg, no workstation sale could mean no MBP sale to that customer. The more niches that Apple doesn't play in, the more it limits how many products they can sell. That Mac Pro enables sales of the MBP because it's keeping a lot of software on the macOS platform for MBP to run.
A gaming PC is a particularly easy and acute example. You do not get a Mac to play games. You get a PC. Well, that means that person may get a Windows laptop instead, or an Android phone instead. There are a lot of apps like this. Not having Creo hurts (the Mac workstation troubles probably directly plays into this). The Matlab implementation is limited. Not having something competitive to CUDA (hopefully yet) and Nvidia's class leading performance hurts. Microsoft doesn't even have all of the Office apps on Mac. Not having MS Project hurts sales of Macs. Who the heck knows what Windows only software that hospitals run? Every niche that Apple doesn't have a foothold limits how many units they can sell.
Okay, TL;DR. Some, though.
What we knew so far is that 4-die Apple Silicon could be 48+16, suggesting a 12+4 config, or a 4*(8+2). To the former it will hit 512GiB, or the latter 256GiB. No way we’re going to see 1.5TiB of RAM, nor does the Mac Pro really have apps that needs them. It’s more of a “flex their muscle because they can”.
Nor we might see the M2 Pro/Max, no rumors I know suggest that’s the plan. I’d bet Apple want their next Pro chips to be M3, which will hit 2023. Good since there won’t be huge competitions next year. For now, it’s M1 (8+2)*2 or 4.
Workstations changed from traditional “Massive CPU, enormous RAM with tons of expansions” towards something more GPU-focused, with decent amount of cores for the job. Anything 32+8 and 48+16 will be sufficient for a long time. As good as Threadripper Pro looks, it’s also a muscle-flex rather than designed for real case scenario. GPUs are going to be massive for sure, with four professional dies, it can rival multi graphics card setup with an unthinkable amount of VRAM. With that said, combined with that level of power efficiency, smaller & cheaper makes sense. That doesn’t mean we’ll go slot-free, but we no longer needs eight of them (one being I/O cards anyway) I’d say somewhere around 4-6, just minus the current MPX modules.
The trash can is a dud also because its power supply, which I believe it bottlenecks the thing, other also being that it doesn’t quite hit the goal they want - that workstations will be relying on more flexible external support rather than internal expansions. It’s good to be more cautious while chasing for that goal. I don’t think they’ll do something that cuts off expansions completely, but shrinking to be just enough is great.
Word about the iMac Pro, so long it exists, it’ll be a tier lower than the Mac Pro. I don’t think we’ll see a 4-die version even if Apple can make one. Just not the target market. People looking for workstations are still looking for more powerful displays with greater expandability, both internal or external. The iMac Pro also needs to be smaller to stay competitive against professional displays, or, what’s the point.
What I think it’ll be a combination of the 5K iMac’s form with Pro’s performance, with 16+4 that’s an easy thing to do. By the way, there are still headroom inside the 16” MacBook Pros, which I think is also plenty for a 2-die. If that’s right, all Apple need is make a larger MacBook Pro on desktops and still kicks 12900K in the arse.
Comments
2013 Mac Pro = 921BTU/h
https://support.apple.com/en-us/HT201796
2017 iMac Pro = 1262BTU/h
https://support.apple.com/en-us/HT208378
2019 Mac Pro = 3076BTU/h
This was with Intel/AMD parts though, which are around 5x less power efficient than Apple Silicon. While they could target the same power level, to do that would need bundling around 10 x M1 Max chips, which would be over 100 TFLOPs of computing power. That seems excessive for what the workstation market needs in a single machine. I expect they will instead target the same performance level at a lower power.
On 3nm, they can get 70% more transistors than 5nm so potentially over 60TFLOPs in under 300W. This is the performance level the next-gen 5nm GPUs are reported to have:
https://wccftech.com/nvidia-ada-gpu-rumored-specs-18432-cores-64-tflops/
Someone could have quadruple those but even today it's rare that people buy more than a single 3090. There a video here where they put 4 x 3090 in an external box:
That's the kind of PCIe box Apple could make for people who need it for non-GPU peripherals. A Mac Pro chip on 3nm would perform roughly like the 2x 3090s in that video. There's also nothing wrong with buying PCs for this kind of task, Apple doesn't have to cater to every use case. Someone can buy a Mac Pro for the creative part and offload the render to a farm of PCs with multiple GPUs.
The 2019 Mac Pro was requested by some of their high profile Mac users:
https://forums.macrumors.com/threads/red-president-played-a-role-in-making-the-new-modular-mac-pro.2059763/
Tim Cook or one of the other execs said that the Mac Pro segment doesn't move the needle for them, it's just a space they decided they wanted to cater to. It works as a halo product. Some of the people who requested this model have businesses that sell PCIe cards so they have a financial incentive for these models to still be sold:
https://www.apple.com/newsroom/2019/06/pro-app-developers-react-to-the-new-mac-pro-and-pro-display-xdr/
Mac Pro buyers only number in the tens of thousands per year but if even 2000 of those buy a $1000 PCIe card, that can make an entire business that sells those cards viable and could cause some businesses to go bankrupt without that support.
The market with infinite computation requirements doesn't have to do it all on a single machine. A single workstation only has to be good enough for real-time tasks like 8K video and offer good performance-per-dollar. Offline computation can be done with multiple machines.
HP is the biggest manufacturer in the workstation market, their 2020 annual report is here:
https://d18rn0p25nwr6d.cloudfront.net/CIK-0000047217/dc81129d-f0cd-4a42-b6e8-46b29044d9b6.pdf
Workstations: $1.8b (2020), $2.3b (2019), $2.2b (2018).
Desktops: $9.8 (2020), $12b (2019), $11b (2018)
Notebooks: $25.7b (2020), $22.9b (2019), $22.5b (2018)
70% notebook, 25% desktop, 5% workstation. For Apple it's closer to 80% notebook, 19% desktop, < 1% workstation.
Globally the unit volume of workstations is reported in different places around 1 million units per quarter of which Dell and HP are the large majority and split evenly, something like 40% each. Assuming HP is selling 1.6m workstations per year, this makes their ASP $1125 per workstation. That means models like these that start around $1200 that have things like Core i5 CPU, 8GB RAM:
https://www.hp.com/us-en/shop/mdp/business-solutions/hp-z2-workstation#!&tab=vao
Those models are recommended by HP for things like architecture planning. The high-end workstation market ($3k+) is a fraction of the overall workstation market and it will get smaller over time the more that people migrate to notebooks, which can now do the jobs workstations were used for.
I'd say the premium segment ($3k+) would likely be no more than 25% of the overall volume so Apple would be competing in a market of around 1m units per year. They won't be able to beat HP/Dell because they are used in the server space but they could possibly get 20% of the premium market at around 200,000 units per year (still < 1% of Mac unit volume). At an ASP of $5k, this is $1b.
Macs last year made $35b, iPhone $191b, iPad $31b, wearables $38b, services $68b. They make way more from even AirPods and the Apple Watch than a best-case for workstations.
If the workstation market decides they prefer external GPUs with add-in boards, Apple would be better just leaving the Intel Mac Pro as is and bump the spec every now and then and put Apple Silicon in the iMac Pro. In about 5 years, there will be no more upgrades to go to and no more demand for premium desktops.
Then, yes, they got feedback that that the 2017 iMac Pro wasn't going to be enough and PCIe slots are an important feature for some of their big customers, however else they arrived at the decision, and the 2019 Mac Pro was a go and they said it will be coming.
If Apple can put 100 TFLOPs in a Mac Pro, at less than 1000 W, they should. People will buy it for their desks. People buy lots of them and rack them. Don't understand the hesitancy here. If Intel, Nvidia or AMD could, they would. You design in performance up to all practical limits for the given form factor. If the half sized Apple Silicon Mac Pro offers 40 TFLOPs at 250 W, well Apple should design in a high bandwidth network for them so people can buy 2, 3, 4 of them on their desks or have lots of them racked. They have to get on with getting HPC codes converted to Metal.
I'm not a proponent of "which came first, the chicken or the egg" type situations btw. The hardware always comes first, and it's always up to the OEM to sell a lot of units with the right customers for developers to target. The developers are always fast followers. So ultimately, Apple has to offer a reasons above and beyond just performance. So, Metal, Metal, Metal conversions of HPC code, which they will likely have to pay for. If they offer a lot of CPU cores, they have to get HPC code optimized for ARM.
There is never enough compute performance in HPC.
I wouldn't make a bet of 5 years. It's the usual push-pull of client vs server computing. Maybe cloud computing will be so cheap, that it will drive high performance workstations down to "not worth Apple's time", but I wouldn't make that bet for 5 years. Wouldn't make it all. There will be applications gated by the network, whether by bandwidth, by security, network stability, or whatnot. So, I think there will always be something that will drive sales of workstations on desks. What was done on a compute cluster yesterday, is done on a workstation today. What is done on a computer cluster today, will be done on a workstation tomorrow. The computer cluster just gets bigger, handling ever bigger problems.
Then, I would say the notion of the "market is too small to be worth Apple's time" as short sighted. Much like the decision to have a 2019 Mac Pro, it wasn't directly about units sold. It's about platform or ecosystem stability. Every niche that Apple doesn't play in is another reason for a loss of a sale of product in their ecosystem. Eg, no workstation sale could mean no MBP sale to that customer. The more niches that Apple doesn't play in, the more it limits how many products they can sell. That Mac Pro enables sales of the MBP because it's keeping a lot of software on the macOS platform for MBP to run.
A gaming PC is a particularly easy and acute example. You do not get a Mac to play games. You get a PC. Well, that means that person may get a Windows laptop instead, or an Android phone instead. There are a lot of apps like this. Not having Creo hurts (the Mac workstation troubles probably directly plays into this). The Matlab implementation is limited. Not having something competitive to CUDA (hopefully yet) and Nvidia's class leading performance hurts. Microsoft doesn't even have all of the Office apps on Mac. Not having MS Project hurts sales of Macs. Who the heck knows what Windows only software that hospitals run? Every niche that Apple doesn't have a foothold limits how many units they can sell.
This affected the 2013 Mac Pro too in that it didn't offer much more than what people already had and the higher volume users at the lower performance levels migrate down to cheaper models. Apple's Mac growth in the last few years has been driven by notebooks. If someone has a specced out 2019 model that already offers up to 64GB video memory, 1.5TB RAM, up to 56TFLOPs GPU, native Windows support for specialist software like CAD, it's hard to make a compelling upgrade in a new Apple Silicon model.
That's why it makes sense to use 3nm. This will give them up to 70% more transistors. This isn't good timing-wise though as 3nm won't be ready next year.
https://www.anandtech.com/show/17013/tsmc-update-3nm-in-q1-2023-3nm-enhanced-in-2024-2nm-in-2025
If M2 will be on 4nm, they can do the iMac Pro early 2022 with M1 Pro up to M1 Max Duo. Later in the year they do M2, M2 Pro, M2 Max for the MacBooks and M2 Max Duo/Quad for the Mac Pro and the iMac Pro can get a refresh. Then it's 3nm for M3 late 2023.
For compute, Nvidia has done a lot software-wise to boost adoption like with raytracing support and other AI libraries to accelerate real-world workflows. Like you say, it would help boost Metal adoption and software performance with more Metal libraries to be able to use all that compute power:
https://developer.nvidia.com/gpu-accelerated-libraries
While I agree that Apple should continue to cater to high-end users so that software support is there, that doesn't have to mean a form factor like the 2019 model. There were pro users that bought iMac Pros but the audience for these high-end models in every form factor is small so every model will seem like a commercial flop. The main thing is they cater to real world workflows and with their work on Apple TV+, they have direct access to higher end visual effects and what the needs are.
It's not just that the Intel Mac Pro that's dead. The entire X86 platform is about to enter assisted living. The platform is decades beyond its expiration date and that company has done nothing, really, to keep their platform modern. And AMD Ryzen also. Soon, all the PC chip makers will be switching over to ARM, and will pretend that they're the innovators. In five or seven years, you'll be hard pressed to find an X86 PC at Best Buy.
I have an M1 Mac mini so as to test and learn about the new tech, but work on my i9 27” iMac 64 GB. The new M1 MBPs are nice but the next logical machine for me will be the 2022 large format iMac unless there is a new Mac Pro, but even if there is I may refrain. If the 2022 iMac has a 6K screen and at least 27” or more and sports a quad M1 Max I would be very, very happy with that, for sure.
What we knew so far is that 4-die Apple Silicon could be 48+16, suggesting a 12+4 config, or a 4*(8+2). To the former it will hit 512GiB, or the latter 256GiB. No way we’re going to see 1.5TiB of RAM, nor does the Mac Pro really have apps that needs them. It’s more of a “flex their muscle because they can”.
Nor we might see the M2 Pro/Max, no rumors I know suggest that’s the plan. I’d bet Apple want their next Pro chips to be M3, which will hit 2023. Good since there won’t be huge competitions next year. For now, it’s M1 (8+2)*2 or 4.
Workstations changed from traditional “Massive CPU, enormous RAM with tons of expansions” towards something more GPU-focused, with decent amount of cores for the job. Anything 32+8 and 48+16 will be sufficient for a long time. As good as Threadripper Pro looks, it’s also a muscle-flex rather than designed for real case scenario. GPUs are going to be massive for sure, with four professional dies, it can rival multi graphics card setup with an unthinkable amount of VRAM. With that said, combined with that level of power efficiency, smaller & cheaper makes sense. That doesn’t mean we’ll go slot-free, but we no longer needs eight of them (one being I/O cards anyway) I’d say somewhere around 4-6, just minus the current MPX modules.
The trash can is a dud also because its power supply, which I believe it bottlenecks the thing, other also being that it doesn’t quite hit the goal they want - that workstations will be relying on more flexible external support rather than internal expansions. It’s good to be more cautious while chasing for that goal. I don’t think they’ll do something that cuts off expansions completely, but shrinking to be just enough is great.
What I think it’ll be a combination of the 5K iMac’s form with Pro’s performance, with 16+4 that’s an easy thing to do. By the way, there are still headroom inside the 16” MacBook Pros, which I think is also plenty for a 2-die. If that’s right, all Apple need is make a larger MacBook Pro on desktops and still kicks 12900K in the arse.