Apple Silicon Mac Pro predictions

[Yebubbleman]

There's an elephant in the room in the Apple Silicon Macs section of forums.macrumors.com. We're pretty sure that the initial Apple Silicon MacBook Air and 13" MacBook Pro models will retain their current/familiar designs, while the Apple Silicon 21.5" iMac replacement will get an iPad Pro inspired redesign and a 24" screen.

But let's turn our attention to something far more nebulous. What do you all think the Apple Silicon replacement to the 2019 Mac Pro (and/or whatever last Intel refresh to it that may or may not be lurking in the interim) will be like? Will it still use AMD cards? Will it use Apple graphics cards? Or will the SoCs used in it take over all graphics leaving any PCIe cards for non-graphics expansion? Will Apple change the form factor up or keep the 2019 Mac Pro design? And what about the future of Afterburner?

What do you think will happen? What would you like to see happen?

(Personally, and especially at the asking price of the current 2019 Mac Pro, I don't have a dog in this fight. The Mac Pro isn't my Mac. Though, it seems like the one with the most nebulous future going into this transition and I'm curious as to how you all feel about it!)

 

[vigilant]

Afterburner will continue to be used, and what we see right now.

I expect Apple will do something similar to what Intel does for their chips. They will produce core dense SOCs, with binned parts to fill out that line. I expect we will see 16 cores for the CPU being the starting line, and harder to say with the GPU considering we don't know necessarily what the roadmap looks like for that.

For ultra high performance products like the iMac Pro, MacBook Pro, and Mac Pro I expect Apple will level up their fabric design to get closer or even surpass AMD's Infinity Fabric.

There isn't really a clean for them to fully take advantage of the dedicated hardware without it.

 

[nothingtoseehere]

At the time Apple made its plans for the Mac Pro, Apple also made its plans for Apple Silicon.
I believe they would not have designed the new Mac Pro just to nix it a few years later when abandoning Intel.
Therefore I think they have plans to use the Mac Pro design and fit it with Apple Silicon chips.

 

[JMacHack]

What I think will happen, based on convos and theories read on these forums:

ASi Mac Pro:
Same look as current, same amount of PCI slots, maybe PCI 4?
"Apple Silicon Pro" CPU, beefed-up ASi chip maybe shared with upper level iMacs/MacBook Pros, will have onboard GPU but not used as main GPU, like dGPU MBPs.
RAM slots, which act secondary to the onboard RAM of the SoC, allowing segmentation to the lineup.
"Apple Pro" GPU, or baseline Radeon GPU, whatever is current at the time of release. If I had to guess it's likely going to be similar to the Afterburner card but with some parts devoted to traditional graphic tasks. Maybe raytracing? 8gb for the baseline, up to 32 (64? I doubt they'd need that much but in 2 years who knows) of GDDR6
I personally don't think they'll lower the upper RAM limit, but others on here said they might, so I'll be conservative and say a max of 1tb of RAM.
Base price of $5000, since the iMac Pro is likely going to be dead and this will fit right above the iMac

What I hope will happen:
ASi Mac Pro
Same look as current, same amount of PCI slots, PCI 4
"Apple Silicon" CPU, same level as high-end iMacs/MacBook Pros, onboard GPU etc. etc.
"Apple Silicon Pro" CPU, bespoke processor for the top-of-the-line Mac Pro, more cores, more cache, more Neural Engine. No iGPU
RAM slots, same as above
"Apple Pro GPU" same as above 8GB GDDR6
"Apple Pro GPU MPX" bigger brother to the Pro GPU, capable of high-end performance 16gb GDDR6 baseline
RAM: 64-2TB (higher amounts require ASi Pro CPU
Base Price of $3000, returning it to the price of previous generations of Mac Desktops, obviously the high end stuff will cost more though.

 

[leman]

I don't think that we have enough information to make educated guesses on this one. My intuition tells me that non-Apple GPUs will have to go — Apple wants to build a predictable GPU programming model and consolidating GPU features is the key. Not to mention that driver quality for third-party hardware is often lacking.

Let's focus on what we do know. We know that Apple has invested a lot of effort in designing a state of the art multi-GPU support. This part of Metal is only applicable to the Mac Pro. I doubt that they would have gone through the trouble for a single hardware iteration. This makes me think that future Mac Pro might have multi-GPU setups connected in complex topologies.

As to how such a system will look in practice, no idea. One thing that puzzles me is Apple's focus on unified memory. Unified memory is great for pro-level workloads — CPU and GPU can access the same memory without having to copy it around through the slow PCI-e interface, which allows some completely new approaches to data processing, with CPU and GPU truly working together (and not like now, where CPU is used to prepare the data and then wait for a while until the GPU sends it back). But I have no idea how unified memory can be reasonably implemented on a scale of a Mac Pro. A single CPU and GPU die could be connected to shared I/O die (just like what AMD is doing these days), but this approach is not practicable in a multi-GPU system. You will need some sort of NUMA setup, with ultra-fast interconnects between the boards....

 

[JMacHack]

But I have no idea how unified memory can be reasonably implemented on a scale of a Mac Pro. A single CPU and GPU die could be connected to shared I/O die (just like what AMD is doing these days), but this approach is not practicable in a multi-GPU system. You will need some sort of NUMA setup, with ultra-fast interconnects between the boards....

I don't think it can. Not when we're talking potential terabytes of RAM anyway. I think the emphasis is on the unified memory for the 1st gen silicon, but the "Pro" machines will have to address more RAM that can be reasonably fit onto a CPU die. Also the whole point of GDDR is that it's better for graphics workloads, so why would they opt for just using the system RAM? I think to at least achieve feature-parity with the current Mac Pro, Apple would need to have at least a dedicated GPU and VRAM. Maybe as an add-in card? They have special MPX slots in the 7,1, maybe they'll be updated with a faster proprietary interconnect.

 

[David Hassholehoff]

At the time Apple made its plans for the Mac Pro, Apple also made its plans for Apple Silicon.
I believe they would not have designed the new Mac Pro just to nix it a few years later when abandoning Intel.
Therefore I think they have plans to use the Mac Pro design and fit it with Apple Silicon chips.

I agree. Apple has been working on Silicon for many years, and the redesign of the Mac Pro makes no sense if it is to be used only for one generation. I expect the first AS based Mac Pros to look exactly like the current cheese grater. As for supporting third party graphics cards? I have no idea, could see it going either way. I hope they will leave it open, though.

Also, I would really love to see something between the Mac Pro and the Mac mini, something like the PM Cube. A Mac mini Pro.

 

[JMacHack]

Also, I would really love to see something between the Mac Pro and the Mac mini, something like the PM Cube. A Mac mini Pro.

I dream of a Mac Mini X, with one PCI slot and a bigger chassis to accommodate. Everyone's always dreamt of the xMac tower and I don't think it's ever gonna happen.

 

[BeatCrazy]

ASi Mac Pro will be the very last Mac to make the transition/update. Remember these are machines used to make money.

 

[Joe The Dragon]

I don't think it can. Not when we're talking potential terabytes of RAM anyway. I think the emphasis is on the unified memory for the 1st gen silicon, but the "Pro" machines will have to address more RAM that can be reasonably fit onto a CPU die. Also the whole point of GDDR is that it's better for graphics workloads, so why would they opt for just using the system RAM? I think to at least achieve feature-parity with the current Mac Pro, Apple would need to have at least a dedicated GPU and VRAM. Maybe as an add-in card? They have special MPX slots in the 7,1, maybe they'll be updated with a faster proprietary interconnect.

But MPX is really just added power and TB and some added pci-e lanes.

now the AMD EPYC dual systems use the pci-e bus for CPU TO CPU links.

 

[vigilant]

I don't think that we have enough information to make educated guesses on this one. My intuition tells me that non-Apple GPUs will have to go — Apple wants to build a predictable GPU programming model and consolidating GPU features is the key. Not to mention that driver quality for third-party hardware is often lacking.

I hear you. One thing to keep in mind that Apple actually writes the drivers for all devices in the Mac. It's part of the reason Apple went away from Nvidia. Nvidia won't let Apple write up to the stack the way they want to.

But I have no idea how unified memory can be reasonably implemented on a scale of a Mac Pro. A single CPU and GPU die could be connected to shared I/O die (just like what AMD is doing these days), but this approach is not practicable in a multi-GPU system. You will need some sort of NUMA setup, with ultra-fast interconnects between the boards....

I 100% agree. What I suspect, is we are going to see a much richer fabric design to allow enhanced access to system memory, and any graphics cards would need to have their own dedicated bank of memory... unless they are able to provide high enough bandwidth through the system for it to get access to memory.... I suspect we'll see a T3 at some point that will be dedicated to provisioning of resources at a far more granular level.

 

[Jorbanead]

At the time Apple made its plans for the Mac Pro, Apple also made its plans for Apple Silicon.
I believe they would not have designed the new Mac Pro just to nix it a few years later when abandoning Intel.
Therefore I think they have plans to use the Mac Pro design and fit it with Apple Silicon chips.

Agreed. Apple is a very foreword thinking company, and Tim Cook always says they have the long game in mind. I see the current Mac Pro form factor to stay for at least the next 5-10 years. I could see them cutting the Mac Pro in half eventually - so you could fit two Mac Pro’s in the same rack space as one, like a Mac Pro “mini” with space for only one MXP module.

My guess is that all AS chips will come with integrated graphics, but you’ll still be able to expand via MPX modules to get even more power. Whether these are AMD or custom AS, who knows. I figure AS GPU’s is the ultimate goal, but Apple will stick with AMD for awhile on the top end MPX modules.

What I’m really curious about is the core count options. My guess is starting out we’ll see a max core count of 36, with 4 efficient cores and 32 performance cores. That, plus something insane like a 128-core neural engine (the A13 already has an 8-core neural engine) for machine learning. Or maybe this could be an MPX add-on.

 

[leman]

One thing to keep in mind that Apple actually writes the drivers for all devices in the Mac. It's part of the reason Apple went away from Nvidia. Nvidia won't let Apple write up to the stack the way they want to.

Are you 100% sure about this? Information I've had so far was conflicting. I had an impression that IHV's have small teams writing drivers for Apple, with Apple's input. As to the conflict with Nvidia, my bet would be that it had to do with CUDA. Apple likely wanted to ban CUDA from their platform to force people in adopting Metal.

Also the whole point of GDDR is that it's better for graphics workloads, so why would they opt for just using the system RAM? I think to at least achieve feature-parity with the current Mac Pro, Apple would need to have at least a dedicated GPU and VRAM.

Depends on what you use for system RAM. Something like HBM would work. GDDR has a lot of bandwidth (can transport a lot of data at once) but very high latency (takes a lot of time to access data), so it's not suitable to feed a CPU. High latency of GDDR might also be an issue for some professional workloads, now that the GPUs are more flexible and memory access is less predictable. There is a reason why HBM is widely used in HPC.

As to Mac Pro, I think what we will see is essentially a hierarchy of caches, with some API to control for cache locality. This is also an architecture that AMD has been emphasizing for a while. Dedicated processors will have their local fast cache, but everything will be backed by some sort of unified memory, with very fast fabric connecting all of that.
Basically, NUMA. As I mentioned, Metal already made first steps towards this programming model. You can discover groups of devices that are "closer" together in terms of memory access and schedule work that needs to address same memory blocks on such groups. I can imagine similar additions on the CPU side, allowing you to pair CPUs and GPUs that can access the same memory without performance penalties.

 

[Boil]

Mac Pro Cube - starting at US$5,999.00

48 P cores / 4 E cores / 96 GPU cores - CPU / GPU Chiplets & RAM on interposer / System in Package (SiP) design
HBM3 Unified Memory Architecture - 128GB / 256GB / 512GB
NVMe RAID 0 (dual NAND blades) 4TB / 8TB / 16TB
Four USB4 (TB3) ports
Four TB4 ports
Two 10Gb Ethernet ports
One HDMI 2.1 port
Three MPX-C slots (for use with asst. MPX-C expansion modules)


Apple MPX-C Expansion Modules - starting at US$599.00

NVMe RAID Storage Module (Quad NAND blades)
GPGPU Module
FPGA Module
Neural Engine Module

 

[vigilant]

Are you 100% sure about this? Information I've had so far was conflicting. I had an impression that IHV's have small teams writing drivers for Apple, with Apple's input. As to the conflict with Nvidia, my bet would be that it had to do with CUDA. Apple likely wanted to ban CUDA from their platform to force people in adopting Metal.

I am 98% certain. Apple writes everything down to the firmware on the cards, to the drivers in macOS, and that is something NVidia simply doesn’t allow it. I say 98% certain because I’m sure theres an odd nuance somewhere in there around something that NVidia would let them do, but it wasn’t enough for Apple.

I had a fairly deep conversation with a friend of mine who works for NVidia as an engineer, not going to name names nor what he does... but yes. It goes to the point where Nintendo Switch full stack for controlling the hardware and boot process is largely all NVidia.

I am not 100% certain on how much of the Nintendo Switch is written by NVidia, but it’s my understanding they contributed to it significantly.

Depends on what you use for system RAM. Something like HBM would work. GDDR has a lot of bandwidth (can transport a lot of data at once) but very low latency (takes a lot of time to access data), so it's not suitable to feed a CPU. Low latency of GDDR might also be an issue for some professional workloads, now that the GPUs are more flexible and memory access is less predictable. There is a reason why HBM is widely used in HPC.

As to Mac Pro, I think what we will see is essentially a hierarchy of caches, with some API to control for cache locality. This is also an architecture that AMD has been emphasizing for a while. Dedicated processors will have their local fast cache, but everything will be backed by some sort of unified memory, with very fast fabric connecting all of that.
Basically, NUMA. As I mentioned, Metal already made first steps towards this programming model. You can discover groups of devices that are "closer" together in terms of memory access and schedule work that needs to address same memory blocks on such groups. I can imagine similar additions on the CPU side, allowing you to pair CPUs and GPUs that can access the same memory without performance penalties.

I really should have read your post before I posted a partial comment.

I like the principles of HBM, but for something like the Mac Pro I’m not sure if you can get to the full customization of a Mac Pro using that alone though. It could easily satisfy the Macbook Pro, MacBook, Mac Mini, iMac and possibly iMac Pro.

 

[t0mat0]

What to make of AMD's Infinity Architecture? AMD would be releasing their 3rd Gen by 2022. Unified memory across CPU and GPU/ Unified data fabric between CPU & GPU. Next gen HBM (HBM 2e?)
More cards you get, more unified memory you have.
Seems like Apple haven't explicitly ruled out AMD as MPX add ons that work with Apple Silicon, much as they do all the Mac Pro GPUs via MPX modules now. Presumably in some capacity, Apple will need to have an interconnect for multi-chip modules ala AMD Epyc
x Apple cores on a SoC die, x SoC dies per Apple MCM

AMD Announces X3D Chip Stacking and Infinity Architecture

AMD takes chip design into the third dimension

www.tomshardware.com

We've seen AMD basically announce things in tandem with Apple product launches before - AMD GPUs in Mac desktops and laptops. Wouldn't be suprised if they synced their announcements to anything they'd do with Apple on any discrete GPUs for Apple Silicon Macs.

We may not hear much till WWDC 2021 beyond what's rumoured for AMD

 

[Kostask]

nVidia won't let anybody write to its hardware, I think they consider it a competitive advantage. Apple won't let anybody write kernel level drivers for MacOS. End result is no nVidia cards on Macs. Not likely to change.

I have heard that the Boot Camp drivers for Mac hardware are written by software houses outside of Apple. That would make sense, as there are a few software houses who can write Windows drivers (as opposed to no software houses who can write MacOS kernel drivers). Apple just licenses those Windows drivers for BootCamp.

 

[leman]

I like the principles of HBM, but for something like the Mac Pro I’m not sure if you can get to the full customization of a Mac Pro using that alone though. It could easily satisfy the Macbook Pro, MacBook, Mac Mini, iMac and possibly iMac Pro.

Imagine add-on cards that contain multiple GPU-like accelerators (maybe even with their own CPU), HBM RAM, all cards connected with very fast interconnect (not PCI-e). Then device affinity APIs that allow you to discover local device groups to maximize performance.

What to make of AMD's Infinity Architecture? AMD would be releasing their 3rd Gen by 2022. Unified memory across CPU and GPU/ Unified data fabric between CPU & GPU. Next gen HBM (HBM 2e?)
More cards you get, more unified memory you have.

Yes, that’s what I mean. I dint think that Apple will use AMD but they might just implement something similar. Apple has been aggressively hiring interconnect engineers for a while now, and they can invest much more into this stuff than AMD ever could.

 

[thenewperson]

Could they go proprietary for all this?

 

[vigilant]

Could they go proprietary for all this?

Possible, but not likely

 

[t0mat0]

The Mac Pro Radeon Pro Vega II MPX Module:

“Infinity Fabric Link connection enables two Vega II MPX Modules to connect at over 80GB/s— 5x the bandwidth of the PCIe x16 gen 3 connection. This connection is installed in the factory when two Vega II MPX Modules are configured, and comes with the Vega II MPX Module when purchased separately.”

The Radeon Pro Vega II Duo MPX Module:
Connected on-board with Infinity Fabric Link

Considering the Mac Pro came out with the world’s most powerful GPU (per FP32)as an exclusive and you can put two in the Mac Pro maybe that’s a tell for the successor to the current Mac Pro.
Was truthfully the Mac Pro a from scratch design when they realised the iMac Pro wouldn’t cut it to be the ultimate Mac years ago? Or did they have the Mac Pro on the schedule for as an Apple Silicon Mac Pro?

“we want to architect it so that we can keep it fresh with regular improvements, and we’re committed to making it our highest-end, high throughput desktop system, designed for our demanding pro customers.

 

[Spock]

I dream of a Mac Mini X, with one PCI slot and a bigger chassis to accommodate. Everyone's always dreamt of the xMac tower and I don't think it's ever gonna happen.

I just can't ever see that happening with the Apple silicon switch.

 

[theorist9]

GDDR has a lot of bandwidth (can transport a lot of data at once) but very low latency (takes a lot of time to access data), so it's not suitable to feed a CPU.

I think you meant high latency, correct?

 

[leman]

I think you meant high latency, correct?

Yes, thank for pointing this out! Fixed my post accordingly.

 

[JMacHack]

I just can't ever see that happening with the Apple silicon switch.

Me neither. Sometimes it's nice to daydream though.