Possible Mac Pro chip options

[MF878]

The rumours are clear that the Mac Pro will likely utilise 2x or 4x M1 Max dies, so I wondered what that could look like in terms of chip and memory options. This info basically guarantees options of 20-core CPU/64-core GPU and 40-core CPU/128-core GPU, as these are 2 or 4 unbinned M1 Max dies. However, it would make sense for Apple to use binning (like on the new MacBook Pro) to increase yield and provide in-between options for a better upsell path.

For example, if Apple utilised the binned 24-core GPU M1 Max, that would create options of:

20-core CPU, 48-core GPU
20-core CPU, 64-core GPU

40-core CPU, 96-core GPU
40-core CPU, 128-core GPU

But what if they went further and binned the CPUs to 8 cores like on the binned 14” M1 Pro? Apple currently doesn’t bin CPU on the M1 Max, so this could be a great way to increase yield while providing a more attainable base option and a middle ground between 20 and 40 CPU cores. Then you’d have:

16-core CPU, 48-core GPU
20-core CPU, 48-core GPU
20-core CPU, 64-core GPU

32-core CPU, 96-core GPU
40-core CPU, 96-core GPU
40-core CPU, 128-core GPU

We know that M1 Max chips come with 32GB or 64GB memory. This means the dual-chip model could be offered with 64GB, 96GB or 128GB. The quad-chip model could also be offered in multiples of 32GB in theory, but it probably makes more sense to simplify to 128GB, 192GB or 256GB. This assumes that Apple would be willing to pair (or quad) M1 Max chips with mismatched memory, and I don’t think they would, so a relatively simple lineup of 64GB (dual only), 128GB and 256GB (quad only) seems more likely to me.

 

[Boil]

If the new Apple silicon Mac Pro lineup is going to launch with the M1 Max SoCs, I would think Apple will be reserving all the top-binned M1 Max SoCs (10-core CPU / 32-core GPU) specifically for the Mac Pro lineup...

And with the possibility of LPDDR5X in the Mac Pro lineup, one could see up to 2TB/s memory bandwidth & up to 1TB RAM...!

 

[CWallace]

I am inclined to agree that the Mac Pro will not have binned versions of the M1 MAX. The whole point of the Mac Pro is "maximum power" and that power comes at a "maximum price" so Apple has no incentive to put binned M1 MAX in the Mac Pro to make it cheaper to purchase.

And if they are not offering binned CPU MAX on the MacBook Pro, there is no way they would start doing so on the Mac Pro.

 

[xraydoc]

It's also possible that Apple is waiting to release the Mac Pro with the M2-generation (M2 Pro / M2 Max) of chips. Not such a stretch if the rumored next-gen MacBook Air has the M2 in it.

 

[CWallace]

It's also possible that Apple is waiting to release the Mac Pro with the M2-generation (M2 Pro / M2 Max) of chips. Not such a stretch if the rumored next-gen MacBook Air has the M2 in it.

That would not track with the Bloomberg rumors and they were 100% accurate with M1 Pro and M1 MAX. And the M1 MAX does have a spare set of components necessary to support running two of them together, which further reinforces the Bloomberg rumor of the 20C/64G base model.

There are rumors of an "M2 Duo" with 16C (8E+8P) and up to 20G, but that might be a special chip for the 24" iMac to offer more performance for those who do not want to go to the (~27") iMac Pro.

At the moment, there are no rumors of an M2 Pro or M2 MAX, but there are rumors of an M3 Pro (Lobos) and M3 MAX (Palma) along with a base M3 (Ibiza) so it might be that the Pro/MAX are on an every-other cycle like the old X models of the A Series for the iPad Pro.

 

[Boil]

It's also possible that Apple is waiting to release the Mac Pro with the M2-generation (M2 Pro / M2 Max) of chips. Not such a stretch if the rumored next-gen MacBook Air has the M2 in it.

April 2022 - new 27" iMac & new taller Mac mini - single & dual M1 Max SoC options
May 2022 - new 13" & 15" MacBook (no more Air) laptops, 24" iMac, & new shorter Mac mini- single M2 SoC
WWDC 2022 - new 30" iMac Pro & Mac Pro Cube - dual & quad M2 Max SoC options
September 2022 - iPad Pro - M2 SoC
October 2022 - 14" & 16" MacBook Pro laptops - M2 Pro/Max SoC options

WWDC 2022 will also bring the last of the Intel Mac Pro models; a final Xeon / RDNA3 space heater

 

[MF878]

I am inclined to agree that the Mac Pro will not have binned versions of the M1 MAX. The whole point of the Mac Pro is "maximum power" and that power comes at a "maximum price" so Apple has no incentive to put binned M1 MAX in the Mac Pro to make it cheaper to purchase.

And if they are not offering binned CPU MAX on the MacBook Pro, there is no way they would start doing so on the Mac Pro.

I would argue the fact that they aren’t offering binned CPU M1 Max options on the MacBook Pro is exactly why they would in the dual/quad configurations.

Currently, if Apple produces an M1 Max that has just one subpar CPU core, I presume they ”chop” (hence the codename Jade C-Chop) this down to an 8-core CPU M1 Pro, which means they‘re wasting up to 16 perfectly good GPU cores + second media engine + second display engine.

However, adding an 8-core CPU M1 Max option to the MacBook Pro would really complicate things. Apple charges $200 to go from 8 CPU/14 GPU to 10 CPU/14GPU. That means an 8 CPU/24 GPU M1 Max would occupy the same price as the 10 CPU/16 GPU, being worse in CPU but better in other areas (GPU, media engine, display engine). The lineup is much simpler without this option, and Apple providing a $200 lower entry point to the M1 Max could negate savings from less wastage.

With dual and quad M1 Max options, you don’t have the issue of overlapping/confusing options. 4 CPU & GPU binned M1 Max chips is still a significantly more powerful system than 2 unbinned chips (32c CPU/96c GPU vs 20c CPU/64c GPU). This would give M1 Max chips with 1 or 2 failed CPU cores somewhere to go where they can be more profitable for Apple, rather than being reduced to a much cheaper M1 Pro.

 

[MF878]

Without binning, all you’d have are two options that are ~$3,000 apart. The quad Max would also require 128GB which I would estimated is an $800 upgrade, so a huge difference between the two unbinned configurations. If we assume this new Mac Pro starts at around $4-5k, that means Apple has to try to persuade you to spend 80-100% more with no middle ground options.

With binned options, Apple hooks you into their game where each step seems like a good deal, and suddenly you’re spending $1-2k more than you planned to. I also think there‘s a very real possibility that a binned dual-chip config could help Apple achieve a start price of $3999. This Mac Pro is going to eliminate many of the things that make the current one very expensive (12x DIMM slots, PCIe slots, 1.4kW PSU, overengineered MPX module). It also can’t look like too crappy a deal compared to the new larger iMac which is now expected to offer a dual M1 Max option, especially if it doesn’t have significantly more upgradeability or expandability.

 

[mr_roboto]

I would argue the fact that they aren’t offering binned CPU M1 Max options on the MacBook Pro is exactly why they would in the dual/quad configurations.

Currently, if Apple produces an M1 Max that has just one subpar CPU core, I presume they ”chop” (hence the codename Jade C-Chop) this down to an 8-core CPU M1 Pro, which means they‘re wasting up to 16 perfectly good GPU cores + second media engine + second display engine.

No, the "chop" refers to something done while designing the masks, not something done in production. In production, they should be running separate wafers and mask sets unique to M1 Pro to make all M1 Pro die.

It's very likely that they just scrap any M1 Max die with a bad CPU core. It's either that, or they're saving them for a future unannounced product.

CPU defects shouldn't be the biggest yield concern with M1 Max. The GPU is a huge fraction of the floorplan; statistically if a Max die has one defect somewhere, it's much more likely to land somewhere in the GPU than anywhere else. This is probably why Apple's only known harvesting option so far is to recover 24-core GPU parts - that's the choice which improves yield the most.

For M1 Pro, which I think it's safe to assume is a far higher volume part, they're harvesting based on both GPU and CPU core failures. GPU failures should still be the most likely, but the half-size GPU reduces the GPU's share of defects, which no doubt makes it more attractive to also harvest CPU core failures.

 

[JouniS]

Currently, if Apple produces an M1 Max that has just one subpar CPU core, I presume they ”chop” (hence the codename Jade C-Chop) this down to an 8-core CPU M1 Pro, which means they‘re wasting up to 16 perfectly good GPU cores + second media engine + second display engine.

I think you are taking binning and chopping too literally.

CPU/GPU cores take a relatively small fraction of the M1 chips. If there is a local defect somewhere, it's quite unlikely that it hits something that can be neatly disabled. It's more likely that binning happens for commercial or quality reasons. Sometimes manufacturers just disable parts of a perfectly functional chip in order to sell it to people who are not willing to pay the full price. And sometimes the chip does not work reliably at full load, but it passes the tests once you disable some parts to reduce power consumption.

 

[MF878]

I think you are taking binning and chopping too literally.

CPU/GPU cores take a relatively small fraction of the M1 chips. If there is a local defect somewhere, it's quite unlikely that it hits something that can be neatly disabled. It's more likely that binning happens for commercial or quality reasons. Sometimes manufacturers just disable parts of a perfectly functional chip in order to sell it to people who are not willing to pay the full price. And sometimes the chip does not work reliably at full load, but it passes the tests once you disable some parts to reduce power consumption.

No, the "chop" refers to something done while designing the masks, not something done in production. In production, they should be running separate wafers and mask sets unique to M1 Pro to make all M1 Pro die.

It's very likely that they just scrap any M1 Max die with a bad CPU core. It's either that, or they're saving them for a future unannounced product.

CPU defects shouldn't be the biggest yield concern with M1 Max. The GPU is a huge fraction of the floorplan; statistically if a Max die has one defect somewhere, it's much more likely to land somewhere in the GPU than anywhere else. This is probably why Apple's only known harvesting option so far is to recover 24-core GPU parts - that's the choice which improves yield the most.

For M1 Pro, which I think it's safe to assume is a far higher volume part, they're harvesting based on both GPU and CPU core failures. GPU failures should still be the most likely, but the half-size GPU reduces the GPU's share of defects, which no doubt makes it more attractive to also harvest CPU core failures.

Yes, I realise now that ”chop” likely refers to the die designs and not a literal process. However, that means they’re throwing away an entire Max chip for 1 defective CPU core, which I think strengthens my point. Obviously the M1 Pro is a higher volume part, but seeing as the M1 Pro and M1 Max use the exact same CPU cores, one would think the CPU failure rate as a proportion of chips produced is practically identical, and therefore if its worth doing on one, then it likely is on the other too rather than discarding all of them. Note that by including the 8-core CPU on the base 14” (likely the volume seller of all the standard configurations), Apple has consigned themselves to selling a large number of binned M1 Pro chips. Also, while the Max is more niche, production of that chip should increase considerably as it makes its way to more models, and is being sold to some customers 2 or 4 at a time.

 

[cmaier]

Yes, I realise now that ”chop” likely refers to the die designs and not a literal process. However, that means they’re throwing away an entire Max chip for 1 defective CPU core, which I think strengthens my point. Obviously the M1 Pro is a higher volume part, but seeing as the M1 Pro and M1 Max use the exact same CPU cores, one would think the CPU failure rate as a proportion of chips produced is practically identical, and therefore if its worth doing on one, then it likely is on the other too rather than discarding all of them. Note that by including the 8-core CPU on the base 14” (likely the volume seller of all the standard configurations), Apple has consigned themselves to selling a large number of binned M1 Pro chips.

The number of die with only 1 bad core and nothing else wrong will be extremely small. The cores are a tiny percentage of the die, and defects are randomly clustered.

 

[MF878]

The number of die with only 1 bad core and nothing else wrong will be extremely small. The cores are a tiny percentage of the die, and defects are randomly clustered.

In that case, would you suppose that the binning of the M1 Pro to 8 CPU cores is mostly artificial, seeing as the base 14” is likely to sell in significant volume?

 

[cmaier]

In that case, would you suppose that the binning of the M1 Pro to 8 CPU cores is mostly artificial, seeing as the base 14” is likely to sell in significant volume?

Yes, likely.

 

[CWallace]

As I understand it from reports showing the actual die, the GPU clusters on the MAX are arranged as two sets of 16 each with the bottom potion of the MAX being only said second set of clusters. So If one of the lower GPU clusters has too many failures (more than 8 and less than 16) they literally "chop" it off and it becomes an M1 Pro with 16 GPU cores.

Of course, this is not the only way to get M1 Pros and there would be masks that only expose the "upper half" to the lithography to make them directly.

 

[LonestarOne]

In that case, would you suppose that the binning of the M1 Pro to 8 CPU cores is mostly artificial, seeing as the base 14” is likely to sell in significant volume?

You’re misunderstanding the way binning works. If a chip has a disabled CPU or GPU core, that does not necessarily mean it has a bad core. It may simply be that the chip is not performing to spec, drawing a bit too much power, and Apple has disabled one core to get it back within acceptable spec. The fact that they disable GPUs more often than CPUs isn’t necessarily “artificial”. It may simply be the best way to get chips back within spec.

Heinlein’s Razor says “Never attribute to malice what can be explained by stupidity.” A corollary of that is never attribute to malice what can be explained by math.

 

[cmaier]

You’re misunderstanding the way binning works. If a chip has a disabled CPU or GPU core, that does not necessarily mean it has a bad core. It may simply be that the chip is not performing to spec, drawing a bit too much power, and Apple has disabled one core to get it back within acceptable spec. The fact that they disable GPUs more often than CPUs isn’t necessarily “artificial”. It may simply be the best way to get chips back within spec.

Heinlein’s Razor says “Never attribute to malice what can be explained by stupidity.” A corollary of that is never attribute to malice what can be explained by math.

It’s not malice to disable functioning cores in order to differentiate product. My former employer did that sort of thing all the time. Have an order for 100,000 chips with X cores, but you only have chips with X+1 cores available? Use fuses to disable the extra cores. Have an order for 200,000 chips with a 2 GHz max speed, but most of your available inventory runs at 2.4 GHz? Use fuses to lock down the clock multiplier.

This happens all the time.

You don’t disable a core because you are out of power spec. Whatever power that core is burning, each of the other cores will burn too (a die is too small for transistor behavior to vary meaningfully across it). So you’d end up with a X-1 core chip that burns too much power compared to your other X-1 core chips.

(Also, “burning too much power” isn’t a thing that happens. You bin based on functionality, first, and then speed. If the chip works, you figure out how fast it works, and then you bin it appropriately.)

 

[Krevnik]

As I understand it from reports showing the actual die, the GPU clusters on the MAX are arranged as two sets of 16 each with the bottom potion of the MAX being only said second set of clusters. So If one of the lower GPU clusters has too many failures (more than 8 and less than 16) they literally "chop" it off and it becomes an M1 Pro with 16 GPU cores.

Of course, this is not the only way to get M1 Pros and there would be masks that only expose the "upper half" to the lithography to make them directly.

Yeah, no (to the bolded part). Chop refers to the floor plan sent to production being chopped down from the Max, not a physical chop of the produced die. Ignoring the feasibility of it for a moment, just consider what yields would be like. The 24-core Max would be easier to produce than a “M1 Pro from M1 Max”, as there’s less of a requirement on where the defect(s) on the die have to land.

It’s much more cost effective yield-wise to be able to produce wafers specifically for the M1 Pro. Let’s run a bit of an experiment using a simulated wafer: https://caly-technologies.com/die-yield-calculator/

Using a 300mm wafer, and assuming that Apple’s dies are square (they aren’t), and using the default defect rates in the calculator (since we don’t have any better estimation and this is just to show a comparison)… the M1 Max would produce 79 good dies and 41 defective ones using these parameters. The M1 Pro would produce 147 with 47 defective. Note, this is best taken as an average you’d expect to see over many dies produced.

Note a couple important things. First, the total defective dies didn’t get much worse, despite nearly doubling the number of good dies. So even if we threw away all the defective dies rather than attempting to salvage them, the smaller die produces nearly double the good chips, and allows me to produce higher quantities faster by using less of the fab’s time. Second, if we assume each wafer is 17000$ (https://www.techspot.com/news/86813-analysts-believe-single-tsmc-5nm-wafer-costs-17000.html) and that the default defect rate in our estimate is reasonable (along with assuming we throw out the defects), then the extra yields mean that my costs per M1 Pro above are around 115$/ea vs 215$/ea for the Max. That’s not insignificant savings. Taking this estimation even further, the M1 itself would be around 39$/ea in the same situation with 434 good dies per wafer. Now, this is not a perfect estimate, but should at least make it very clear that cutting down Max dies after the fact is not even remotely efficient.

 

[MF878]

You’re misunderstanding the way binning works. If a chip has a disabled CPU or GPU core, that does not necessarily mean it has a bad core. It may simply be that the chip is not performing to spec, drawing a bit too much power, and Apple has disabled one core to get it back within acceptable spec. The fact that they disable GPUs more often than CPUs isn’t necessarily “artificial”. It may simply be the best way to get chips back within spec.

Heinlein’s Razor says “Never attribute to malice what can be explained by stupidity.” A corollary of that is never attribute to malice what can be explained by math.

I don’t think I’m misunderstanding, it’s that those specific details aren’t really relevant to the points I was making so I’m not sure why they need mentioning? Yes, I’m sure there are other reasons to disable cores other than simply “this core doesn’t work”, but that doesn’t really change my overarching point which is that if it is worth doing on the M1 Pro (for reasons of yield and having a more comprehensive lineup that caters to more price points), then there’s no reason to suggest it isn’t also worth doing on the Max.