Why a $1000 upgrade from 48 to 64 core M1 Ultra, when two M1 Max upgrades cost $400?

[Fomalhaut]

I'm curious why there is such a big jump (US$1000) from the 48-core GPU M1 Ultra to the 64-core version?

On the M1 Max a 24 to 32 core GPU upgrade costs $200, so you would think that the Ultra upgrade would be twice this ($400) given the fact it is essentially two M1 Maxes.

AFAIK, everything else on the 48-core and 64-core M1 Ultra versions is the same (memory, neural engine, CPU cores), so you are just paying for the GPU core upgrade.

What am I missing?

The $2000 jump from the M1 Max to M1 Ultra also seems hard to justify. It implies the upgrade cost of the SoC is the same as the entire base M1 Max version - i.e. the entire cost of the computer is the M1 Max SoC, so having "two of them" doubles the price of the machine. This is clearly ridiculous - I would be surprised if the M1 Max SoC was much more than half the cost of the whole base-level machine.

The M1 Ultra doesn't appear to be great bang-for-buck - it isn't "twice as powerful" as the M1 Max, but looks like it costs about 3 times as much (if we assume $1000 for the M1 Max SoC). I understand that the laws of diminishing returns apply here, but the pricing looks a bit off to me.

Could the M1 Ultra really have much higher fabrication costs than two M1 Maxes?

I suppose that if you really need this performance, you will pay the price, but it appears to be poor value.

 

[crazy dave]

I think the jump from the Max to the Ultra is justified as the Ultra is a more niche product (requires higher margins) and to be honest $4000 is less than what I thought the base 20 core machine would come in at (although I thought they might offer a cut down 16-core for less than $4000). Of course I also estimated that the base M1 Max machine would be more expensive (though I thought it might come with limited internal expansion like a single x8 PCIe slot or equivalent). Also, the base Ultra is 1TB SSD vs 512 GB and 64GB RAM vs 32GB. So it's really $1400 to go from 10 to 20 cores in CPU and 24 to 48 cores in GPU with the same RAM/SSD. That's reasonable (to me). The base M1 Max studio is a very good value and the base M1 Ultra is good, solid value.

However, the 48-to-64 core jump. That I have to agree: $1000 for that seems excessive (to me). I would've though $600 or so, $800 max (again factoring the need for higher margins at higher product tiers). $1000 is ... hmmm ... pricey, very pricey.

 

[Fomalhaut]

I think the jump from the Max to the Ultra is justified as the Ultra is a more niche product (requires higher margins) and to be honest $4000 is less than what I thought the base 20 core machine would come in at (although I thought they might offer a cut down 16-core for less than $4000). Of course I also estimated that the base M1 Max machine would be more expensive (though I thought it might come with limited internal expansion like a single x8 PCIe slot or equivalent). Also, the base Ultra is 1TB SSD vs 512 GB and 64GB RAM vs 32GB. So it's really $1400 to go from 10 to 20 cores in CPU and 24 to 48 cores in GPU with the same RAM/SSD. That's reasonable (to me). The base M1 Max studio is a very good value and the base M1 Ultra is good, solid value.

However, the 48-to-64 core jump. That I have to agree: $1000 for that seems excessive (to me). I would've though $600 or so, $800 max (again factoring the need for higher margins at higher product tiers). $1000 is ... hmmm ... pricey, very pricey.

Good point considering the increase in SSD and RAM between the base model for Max and Ultra, so as you say it's a $1400 upgrade for the Ultra SoC. However, it's a $2400 upgrade from M1 Max (24-core) to the 64-core GPU Ultra, so still $1000 hike for 16 GPU cores.

Obviously there is quite a high cost associated with the creation of the two-die Ultra that Apple is trying to recoup from people with deep enough pockets to go for the top-tier specs.

 

[crazy dave]

Good point considering the increase in SSD and RAM between the base model for Max and Ultra, so as you say it's a $1400 upgrade for the Ultra SoC. However, it's a $2400 upgrade from M1 Max (24-core) to the 64-core GPU Ultra, so still $1000 hike for 16 GPU cores.

Obviously there is quite a high cost associated with the creation of the two-die Ultra that Apple is trying to recoup from people with deep enough pockets to go for the top-tier specs.

Yeah maybe that's the internal justification. Although I do have to say given that I thought the base models would be more expensive I suppose it all evens out. $5000 for the 20/64 CPU/GPU is roughly where I thought it would be.

 

[russell_314]

It’s called if you really need 64 cores then you will pay what it costs. There’s going to be two types of people buying this machine. One is someone with a lot of money and is probably going use it to check their Facebook and the other is someone using it in a pro setting. The person using it for work doesn’t care about the cost because it makes him or her money and the person that’s using it to check their Facebook already has so much money that they don’t care about the cost

 

[UBS28]

nvm

 

[ReanimationLP]

Price might be partially because if anything goes wrong in the packaging process to put the two processor dies on the substrate together, they're both most likely worthless.

 

[repoman27]

Just a guess, but for a M1 Ultra with a 64-core GPU, in addition to two fully functional M1 Max dies, Apple also has to bin heavily based on parametric yields to get 64 cores that can all operate at the required clocks within the thermal constraints of a single package. The silicon interposer likely requires the power draw of the two chips to be fairly evenly distributed as well. And the inter-chip interconnects on the M1 Max are fused off, so those don't need to be functional.

So I'm chalking it up to the M1 Ultra with 64-core GPU requiring two perfect matched dies with extremely low leakage. Or Apple just thinks that the type of people who want those extra cores are willing to pay that much for them.

 

[Fomalhaut]

Just a guess, but for a M1 Ultra with a 64-core GPU, in addition to two fully functional M1 Max dies, Apple also has to bin heavily based on parametric yields to get 64 cores that can all operate at the required clocks within the thermal constraints of a single package. The silicon interposer likely requires the power draw of the two chips to be fairly evenly distributed as well. And the inter-chip interconnects on the M1 Max are fused off, so those don't need to be functional.

So I'm chalking it up to the M1 Ultra with 64-core GPU requiring two perfect matched dies with extremely low leakage. Or Apple just thinks that the type of people who want those extra cores are willing to pay that much for them.

Interesting idea - I tend to suspect it is the latter of these two possibilities!

 

[bobmans]

Most likely related to yields. For the Ultra they probably need 2 perfect M1 Max’s and then basically glue them together lowering yields even more. Is it a $600 difference? Probably not, but this is still Apple and they’ll sell every single studio they can build for the next couple of months, even at this price.

 

[LonestarOne]

Yeah maybe that's the internal justification. Although I do have to say given that I thought the base models would be more expensive I suppose it all evens out. $5000 for the 20/64 CPU/GPU is roughly where I thought it would be.

And the Intel Mac Pro *starts* at $6000.

 

[MRxROBOT]

NVM misread the post.

 

[ader42]

It’s going to be scarcity of fully functional 32 core chips, pure and simple.

MBP with 24 core M1 Max are much more readily available to buy for example.

 

[maflynn]

I'm curious why there is such a big jump (US$1000) from the 48-core GPU M1 Ultra to the 64-core version?

Because they can.

I'm not trying to sound like a jerk, its clearly a business decision that arrived too based on a number of factors that we are not privy too. Could be that low success rate of producing 64 core versions, could be added complexity, could be that apple just wants better profit margins.

 

[dieselm]

The customer paying $200 to go to 32 cores from 24 cores would likely have been happy with 24. Free profit margin.

The customer paying $1000 to go to 64 cores from 48 cores knows they need it and isn't sensitive to price. Free profit margin.

 

[theorist9]

Just a guess, but for a M1 Ultra with a 64-core GPU, in addition to two fully functional M1 Max dies, Apple also has to bin heavily based on parametric yields to get 64 cores that can all operate at the required clocks within the thermal constraints of a single package. The silicon interposer likely requires the power draw of the two chips to be fairly evenly distributed as well. And the inter-chip interconnects on the M1 Max are fused off, so those don't need to be functional.

So I'm chalking it up to the M1 Ultra with 64-core GPU requiring two perfect matched dies with extremely low leakage. Or Apple just thinks that the type of people who want those extra cores are willing to pay that much for them.

I don't know if this is true, but it certainly sounds like a reasonable guess.

And, consistent with this, the US wait times on Apple.com for the Ultra Studio (2 months base/3 months BTO) are much longer than for the Max Studio (2 weeks base/1 month BTO)

 

[bobcomer]

What am I missing?

Having those 2 M1 Maxes hooked together by an interconnect and functioning as one processor. Not an easy task.

 

[pi=e=3]

You’re paying for yield.

An ultra is a massive chip. Ergo, the chance of a fault existing in its production is higher. More chips end up being recycled, and thus you are paying for that.

The max in contrast has a higher yield, they don’t recycle as many.

 

[Apple Knowledge Navigator]

I'm curious why there is such a big jump (US$1000) from the 48-core GPU M1 Ultra to the 64-core version?

My guess is that it's to do with yields. The full 32-core Max chips will undoubtedly be 'rarer' in production - keep in mind that it only takes one bad core and then you have a 31-core Max... which Apple doesn't sell, so some of the remaining cores are disabled and it must become a 24-core version.

 

[leman]

Because Apple can, that’s why. The prices are not formed using the manufacturing cost + some markup, it’s much more subtle than that. And it’s a fairly standard practice in the industry. Like Nvidia charging 5-10x more for their professional GPUs that use the same chip as the gamer GPU with marginally more pricey RAM. It doesn’t have to make sense logically. You price it at the level people are willing to pay for. Makes much more sense. Why leave money on the fooor?

 

[QuietOC]

Most likely related to yields. For the Ultra they probably need 2 perfect M1 Max’s and then basically glue them together lowering yields even more.

More likely they actually cut Ultras apart to make Maxes. The interconnect is there on the Max dies.

 

[mr_roboto]

More likely they actually cut Ultras apart to make Maxes. The interconnect is there on the Max dies.

If they cut ultras apart to make Maxes, they wouldn't need the special interconnect block at all, just a safe zone where it's OK to cut through a bunch of wires. They need that strip of interconnect for a 2D grid of connection points to a bridge device which connects to the other Max die.

 

[leman]

More likely they actually cut Ultras apart to make Maxes. The interconnect is there on the Max dies.

Nope. Ultras are built from two Max dies. Look up “chiplets”. Producing a large die like the Ultra in one would be commercially unfeasible at this point.

 

[altaic]

If they cut ultras apart to make Maxes, they wouldn't need the special interconnect block at all, just a safe zone where it's OK to cut through a bunch of wires. They need that strip of interconnect for a 2D grid of connection points to a bridge device which connects to the other Max die.

I don’t suppose that anyone has publicly shared a die shot of the ultra? I’m eagerly awaiting having a look.

 

[BanditoB]

I don’t suppose that anyone has publicly shared a die shot of the ultra? I’m eagerly awaiting having a look.