Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

Boil

macrumors 68040
Oct 23, 2018
3,477
3,172
Stargate Command
To be small enough to fit into a Cube (we will even say slightly larger than the original G4 Cube, use the current Mac mini as a footprint & scale it up in the vertical). When you get to the size of PCB that would fit in a modern Cube, there is only room for the PCIe slot, not the MPX slot behind it. So I am afraid MiniMPX cannot be a thing.

If an end user wants expansion via PCIe / MPX slots, they will need to go with the Mac "Big Chungus" Pro chassis. iMac Pros & Mac Pro Cubes will use HPC APUs, with expansion via USB4 ports.

I could see a Mac Pro Cube with a single internal expansion slot, intended for an Apple Silicon compute / render GPU. A/V DSPs & I/O could be handled via external units off of USB4. Could make a really powerful DCC workstation in a compact desktop package.

Please make it in Space Grey & offer a matching monitor (at a reasonable price).

APUzilla - Small - 32 P cores / 4 E cores / 48 GPU cores / 16GB HBM2e UMA
128GB DDR5 RAM
4TB (2@2TB blades) NVMe SSD
600W Platinum-rated PSU
Apple Silicon Compute / Render card - Medium - 64 GPU cores / 24GB HBM2e
32" 6K3K Ultrawide Apple Monitor (built-in speakers, microphone, & webcam)
Apple Low-Profile Mechanical Keyboard
Apple Magic Mouse 3D
2 year AppleCare on Cube & Monitor

US$9,999.00


Other APUzillas:

Small - 32 P cores / 4 E cores / 48 GPU cores / 16GB HBM2e UMA
Medium - 48 P cores / 6 E cores / 64 GPU cores / 32GB HBM2e UMA
Large - 64 P cores / 8 E cores / 80 GPU cores / 64GB HBM2e


Apple Silicon Compute / Render cards:

Small - 48 GPU cores / 16GB HBM2e
Medium - 64 GPU cores / 24GB HBM2e
Large - 80 GPU cores / 32GB HBM2e


Yeah, all A/V expansion (DSPs & I/Os) via USB4, a secondary GPU that is compute / render specific (no display output); kinda like the trashcan Mac Pro...!

I am all over the place with Tile Memory amounts...!

Can we not bring the cube back? I know it's fun to look back at it nostalgically but we need to remember that it was a market failure. As was the Mac Pro 6,1. Gorgeous design, wrong implementation.

But this is fun so I'll have a go:
Mac Pro 8,1:

Base:
• 16-core, 16-thread cpu at 3.0GHz Turbo up to 4.0GHz. Insane amounts of cache, much bigger space dedicated to the Neural Engine
• MPX AMD Radeon Pro (whatever the designation is by then) el cheapo GPU, 2x Thunderbolt out, 2x HDMI out 16gb GDDR6 RAM
OR
• MPX Apple Pro GPU, 4x Thunderbolt out, Afterburner ASICs on board. 16gb GDDR6 RAM
• 512gb storage
• 64gb DDR5 ECC RAM (up to 1tb)
• 7 PCI-E 4 slots
• 1400 watt PSU
• $5000 (maybe $4,000 depending)

Mid:
• 64-core, 64-thread cpu at 3.2Ghz turbo up to 4.3GHz. Even Insanely-er amounts of cache, enormous amount of space for Neural Engine
• MPX AMD Radeon Pro (whatever the designation is) decent-tier GPU 16gb HBM2 RAM same output
OR
• MPX Apple Even-More-Pro GPU, same output, Afterburner, 32gb HBM2 RAM
• 3tb storage
• 512gb DDR5 ECC Ram (up to 1.5 tb)
• yadda yadda same as above
• $9,000-$10,000

Max:
• 128-core, 128-thread cpu at 2.8GHz boost up to 3.8GHz. Xbox hueg amount of Cache, Big thinky Neural engine
• 2x MPX AMD Radeon Pro top-tier GPU 64gb HBM2 RAM
OR
• MPX Apple Pro-est GPU, Afterburner, 64gb HBM2 RAM
• 10tb storage
• 3tb DDR5 ECC Ram (max'd out)
• same as above
• If you have to ask, you can't afford it.

See my self-quote above...!

I really think Apple is going to move away from AMD graphics. They have been slowly moving everything towards Metal (dropping OpenGL, dropping OpenCL, etc.) & AMD was the vendor that 'went thru it' to transition to Metal in their Apple drivers, they needed the influx of cash from Apple (no matter how small, AMD needs R&D funds); Nvidia said "nope, we got our CUDA thing here, we have no interest in integrating Metal in our Mac drivers"...?

We will see ini a few years, yeah...?!? ;^p

Looks good. I think this CPU definitely calls for a Threadripper-style chiplet design. You can put 16 Firestorm cores on one die and surround an on-package I/O die with 1 or 4 or 8 of them.

How do you envision the Apple GPU looking? The Pro Vega II Duo is actually also a chiplet design with two GPU dies on the package... is that what you would envision for Apple?

I really hope Apple does create some kind of multidie beast soon because I want to know what their high-speed interposer solution looks like.

Pro Vega II Duo is two separate packages on one PCB, one die per package...

I would see Apple Silicon GPUs as part of the controlling APU. Apple Silicon GPGPUs would be a similar package, just with a focus on the GPU cores...?
 

awesomedeluxe

macrumors 6502
Original poster
Jun 29, 2009
262
105
Pro Vega II Duo is two separate packages on one PCB, one die per package...

I would see Apple Silicon GPUs as part of the controlling APU. Apple Silicon GPGPUs would be a similar package, just with a focus on the GPU cores...?
My bad. I'm less knowledgeable about the desktop parts.

I'm not quite sure what you're recommending here - Apple GPUs as part of the APU I understand, that's the baseline for Apple Silicon. GPGPUs in a similar package with more GPU cores - are you just saying you want a bigger APU with more graphics cores and the graphics cores doing more things, or is this in addition to the first APU?

If the former, I don't think the GPU needs to be a totally separate part, but I'm not sure I want my primary GPU on the same die as my, uh, 64 Firestorm cores (wait, I don't even want 64 firestorm cores on the same die). If we're willing to cut that core count in half I'm happy to have my GPU on the same package, separated by an I/O die with 64GB shared HBM2E.

EDIT: I saw your post outlining what you have in mind in another thread. So basically you want to keep the APU an APU, and connect out to 2x2 96 core GPGPU modules. That's just way more power than I am able to wrap my head around.

What is the upside to hanging on to the graphics cores in the APU if you've got 384 graphics cores elsewhere? I ask because I think the CPU cores will want a lot more power than any individual GPU core and keeping 96 GPU cores nearby is going to force the CPU to throttle down a bit.
 
Last edited:

Boil

macrumors 68040
Oct 23, 2018
3,477
3,172
Stargate Command
My bad. I'm less knowledgeable about the desktop parts.

I'm not quite sure what you're recommending here - Apple GPUs as part of the APU I understand, that's the baseline for Apple Silicon. GPGPUs in a similar package with more GPU cores - are you just saying you want a bigger APU with more graphics cores and the graphics cores doing more things, or is this in addition to the first APU?

If the former, I don't think the GPU needs to be a totally separate part, but I'm not sure I want my primary GPU on the same die as my, uh, 64 Firestorm cores (wait, I don't even want 64 firestorm cores on the same die). If we're willing to cut that core count in half I'm happy to have my GPU on the same package, separated by an I/O die with 64GB shared HBM2E.

EDIT: I saw your post outlining what you have in mind in another thread. So basically you want to keep the APU an APU, and connect out to 2x2 96 core GPGPU modules. That's just way more power than I am able to wrap my head around.

What is the upside to hanging on to the graphics cores in the APU if you've got 384 graphics cores elsewhere? I ask because I think the CPU cores will want a lot more power than any individual GPU core and keeping 96 GPU cores nearby is going to force the CPU to throttle down a bit.

Xeon W CPUs are advertised at 225W, pinning the CPU usage will go past that...

Radeon Pro Vega IIs are also advertised at 225W (?)...

Point being, I feel a high-performance high-core-count Apple Silicon APU would be hard-pressed to hit 225W with all cores (CPU & GPU) pegged out, I would expect about 180W maximum?

You will note I list the add-in cards as GPGPUs, as they will most likely (from a "Pro user" perspective) be used for GPU rendering, or video processing (DaVinci Resolve loves more GPU cores to work with), stuff like that.

I would expect a hypothetical "2x2 96 core" GPGPU to pull no more than a single Vega II, so about 225W max...?

Time will tell, but I think Apple is really going to surprise everyone in regards to what an arm64-based Apple Silicon Mac Pro line-up can be...!
 
  • Like
Reactions: awesomedeluxe

JMacHack

Suspended
Mar 16, 2017
1,965
2,424
I am all over the place with Tile Memory amounts...!



See my self-quote above...!

I really think Apple is going to move away from AMD graphics. They have been slowly moving everything towards Metal (dropping OpenGL, dropping OpenCL, etc.) & AMD was the vendor that 'went thru it' to transition to Metal in their Apple drivers, they needed the influx of cash from Apple (no matter how small, AMD needs R&D funds); Nvidia said "nope, we got our CUDA thing here, we have no interest in integrating Metal in our Mac drivers"...?

We will see ini a few years, yeah...?!? ;^p



Pro Vega II Duo is two separate packages on one PCB, one die per package...

I would see Apple Silicon GPUs as part of the controlling APU. Apple Silicon GPGPUs would be a similar package, just with a focus on the GPU cores...?
I'd like to see Apple go with the "separate dies on an interposer" system that AMD pioneered with the Zen architecture too. It's just too good to ignore, however I'm not sure how restrictive the patents are and if they can do that.

Looks good. I think this CPU definitely calls for a Threadripper-style chiplet design. You can put 16 Firestorm cores on one die and surround an on-package I/O die with 1 or 4 or 8 of them.

How do you envision the Apple GPU looking? The Pro Vega II Duo is actually also a chiplet design with two GPU dies on the package... is that what you would envision for Apple?

I really hope Apple does create some kind of multidie beast soon because I want to know what their high-speed interposer solution looks like.
I would envision the "Apple GPU" being similar to however their current integrated GPUs are on their CPUs. Although obviously with more cores, and have the afterburner ASICs and Video Codecs built into it. Basically splitting their CPU off into two different dies and using the extra space from the split to stuff more cores into it.

What I'm not sure about is whether it'd be soldered(socketed maybe? Tough to tell since the current 7,1 has a socketed cpu) to the board or on a PCI card like the Afterburner and current GPU cards. I'd personally like to see them socketed or as expansion cards, but if the whole CPU lineup is made with the "solder ball" manufacturing in mind I can't see apple setting up a separate line with contact pins.

Realistically, I'd guess that the CPU would at least be soldered, in the same way the T2 chip is. Assuming of course that they roll the T2's features onto the CPU itself.

I'm convinced that multi-die designs on interposers are the way to go after watching this video.
 
  • Love
Reactions: awesomedeluxe

awesomedeluxe

macrumors 6502
Original poster
Jun 29, 2009
262
105
I'd like to see Apple go with the "separate dies on an interposer" system that AMD pioneered with the Zen architecture too. It's just too good to ignore, however I'm not sure how restrictive the patents are and if they can do that.


I would envision the "Apple GPU" being similar to however their current integrated GPUs are on their CPUs. Although obviously with more cores, and have the afterburner ASICs and Video Codecs built into it. Basically splitting their CPU off into two different dies and using the extra space from the split to stuff more cores into it.

What I'm not sure about is whether it'd be soldered(socketed maybe? Tough to tell since the current 7,1 has a socketed cpu) to the board or on a PCI card like the Afterburner and current GPU cards. I'd personally like to see them socketed or as expansion cards, but if the whole CPU lineup is made with the "solder ball" manufacturing in mind I can't see apple setting up a separate line with contact pins.

Realistically, I'd guess that the CPU would at least be soldered, in the same way the T2 chip is. Assuming of course that they roll the T2's features onto the CPU itself.

I'm convinced that multi-die designs on interposers are the way to go after watching this video.
I loved that video. Thanks for sharing it! I have been wondering for some time if Apple does have a high-speed interposer solution. This would be huge for the Mac Pro, which could connect four 8-core firestorm dies with the misaligned butterfly donut design.
 

awesomedeluxe

macrumors 6502
Original poster
Jun 29, 2009
262
105
Slimmer Macbook Pro 16" ...with a Mechanical Keyboard!

Here's another wild idea. Could the increased transistor density and decreased redundancy of silicon make room for... a mechanical keyboard? I think it could! The key parts of this machine are made on TSMC's N5P process, and it would release 2H 2021.

The slimmest mechanical switch I know of that uses an actual spring design (no scissor!) is the Kailh Choc Mini. It's about .8mm altogether. Here's a shot of them from anandtech:
1594926256010.png

So including backlighting and our bottom case, the keyboard part of the machine would be 1cm thick. And that's fine as long as there is nothing under it. Including the screen (which is mini led and touch because why not), our machine will be about 1.3cm thick.

The goal will be to stick all of our silicon around the keyboard, in a ﹇ bracket shape. To do this we'll make a little room by shrinking our touchpad and pushing down our keyboard just a tad. We need 5cm between the top of the number keys and the top of the base (we can still put silicon under the touchbar, which is suddenly a blessing).

Right now, the silicon in the MBP 16 looks like this (courtesy iFixit):

1594925437377.png


But a lot of that can be made smaller or eliminated. For instance, we don't need both the orange (RAM) and green (GDDR) boxes if we're using two stacks of unified HBM2E, which would be smaller than the green box. We don't need the blue box (Apple coprocessor) if everything it does is integrated into the APU. And the yellow box (AMD GPU) and red box (Intel CPU) can both be much smaller. We're going to unify all five of these parts on one 1.5cm (H) x 3cm (W) silicon interposer, which goes above the keyboard and under the touchbar. On top of the interposer is, from left to right:

APU Die (A14Z)
8x Perf Cores ~3.2GHz
4x Efficiency Cores ~2GHz
12x GPU Cores

Memory Die
2x HBM2E Stacks = 32GB

GPU Die (Bionic 14-28)
28x GPU cores

Power and Performance: The Perf Cores are clocked 20% faster than the A13, relying both on the improvements of the N5P process and some additional power (about 2W / core). They're based on the A14 design, but have a little more cache and clock. The GPU Cores are clocked 15% faster than the A13, relying entirely on N5P's process improvements (about 1W / core). They cores use the A14 Bionic design with no changes except the clock boost from N5 -> N5P.

The total TDP of this thing, including 10W for the two stacks of HBM and 2W for the efficiency cores, is around 56W. Accordingly, it is flanked by two small (20mm diameter) fans, which I think will cool it. It's fortunately positioned, being able to vent up out of the base or down under the roomy mechanical keys.

The base configuration uses a 24-core GPU and just one stack of HBM. It's the same part, but salvages some of the semidefective GPU dies. Battery life should be incredible - there's still plenty of room for battery under the lower half of the system, and we're only using about 2/3 the power of a current gen MBP. I think performance will be much better too - I'm confident about the CPU, but it's harder to guess how a 28-core Apple GPU die would compare to a Radeon Pro 5600M. Still starts at $2400... but the extra GPU cores and HBM2E are $700 a piece, the touchscreen is $400, and the SSD price gouging is insane ;)
 
  • Like
Reactions: Boil

Boil

macrumors 68040
Oct 23, 2018
3,477
3,172
Stargate Command
Slimmer Macbook Pro 16" ...with a Mechanical Keyboard!

APU Die (A14Z)

8x Perf Cores ~3.2GHz
4x Efficiency Cores ~2GHz
12x GPU Cores

Memory Die
2x HBM2E Stacks = 32GB

GPU Die (Bionic 14-28)
28x GPU cores

Power and Performance: The Perf Cores are clocked 20% faster than the A13, relying both on the improvements of the N5P process and some additional power (about 2W / core). They're based on the A14 design, but have a little more cache and clock. The GPU Cores are clocked 15% faster than the A13, relying entirely on N5P's process improvements (about 1W / core). They cores use the A14 Bionic design with no changes except the clock boost from N5 -> N5P.

The total TDP of this thing, including 10W for the two stacks of HBM and 2W for the efficiency cores, is around 56W. Accordingly, it is flanked by two small (20mm diameter) fans, which I think will cool it. It's fortunately positioned, being able to vent up out of the base or down under the roomy mechanical keys.

The base configuration uses a 24-core GPU and just one stack of HBM. It's the same part, but salvages some of the semidefective GPU dies. Battery life should be incredible - there's still plenty of room for battery under the lower half of the system, and we're only using about 2/3 the power of a current gen MBP. I think performance will be much better too - I'm confident about the CPU, but it's harder to guess how a 28-core Apple GPU die would compare to a Radeon Pro 5600M. Still starts at $2400... but the extra GPU cores and HBM2E are $700 a piece, the touchscreen is $400, and the SSD price gouging is insane ;)

So the ASi version of KabyLakeG? ;^p

J/K, I like it.

And I have been wishing for Apple to make a low-profile mechanical keyboard for the Mac line-up forever now. Like my HexGears X1 I am typing this on, but Aple-ified!
 

awesomedeluxe

macrumors 6502
Original poster
Jun 29, 2009
262
105
So the ASi version of KabyLakeG? ;^p

J/K, I like it.

And I have been wishing for Apple to make a low-profile mechanical keyboard for the Mac line-up forever now. Like my HexGears X1 I am typing this on, but Aple-ified!
That pretty much is what it is though, if I had to describe it in one sentence. I kind of think separating the GPU cores and the CPU cores is best for both of them. Though also, two stacks of HBM2E is 10W in a very dense place and maybe not the best barrier. But I can't do a pinout for reasons.

I know it's impossible, but a mechanical keyboard on an MBP would be cool, right?
 
  • Like
Reactions: Boil
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.