M5 Studio specs?

[chars1ub0w]

From a Taiwan newsletter (工商時報; Commercial Times) by a Mr. Chang (張), perhaps these are the specs of the M5 Studio?
36 CPU cores, 84 GPU cores and 512GB Max Unified RAM. More than double the M5 Max GPU cores, but less than double the CPU cores.

 

[osx64]

M5 Max MacBook Pros have 18 CPU cores, not 16 or 24.

 

[Pressure]

I also don't see how it can have a 84-core GPU 😂

 

[chars1ub0w]

I also don't see how it can have a 84-core GPU 😂

Why not? "... with the new SoIC-MH 2.5D technology from TSMC, Apple has separated the CPU/NPU from the GPU, moving away from the monolithic design era that limited scaling opportunities for M-series SoCs. Now, Apple can independently scale the size of the CPU clusters with more cores and scale the GPU die with many more cores."

 

[Pressure]

Why not? "... with the new SoIC-MH 2.5D technology from TSMC, Apple has separated the CPU/NPU from the GPU, moving away from the monolithic design era that limited scaling opportunities for M-series SoCs. Now, Apple can independently scale the size of the CPU clusters with more cores and scale the GPU die with many more cores."

And you think that would logically limit them to an arbitrary 4 more GPU cores as opposed to the old "double up"?

We are simply saying these rumors don't even pass the smell test but I guess we will see 🤷🏼‍♂️

 

[chars1ub0w]

And you think that would logically limit them to an arbitrary 4 more GPU cores as opposed to the old "double up"?

I don't know. But if there's is this freedom not to simply double up, Apple surely will look at the target market and try to make the CPU/GPU mix as attractive to as many people as possible. The fact that a weird number is being reported means either Mr. Chang doesn't know what he's talking about or Mr. Chang is well aware of SoIC-mH and has inside information,

 

[Pressure]

I don't know. But if there's is this freedom not to simply double up, Apple surely will look at the target market and try to make the CPU/GPU mix as attractive to as many people as possible. The fact that a weird number is being reported means either Mr. Chang doesn't know what he's talking about or Mr. Chang is well aware of SoIC-mH and has inside information,

I'll acknowledge that it appears we are at opposite sides of that argument with regards to that Chang fella 😂

I don't think it makes any sense for Apple to produce a 84-core GPU "tile" that will only ever be used in their lowest volume computer sold. It would also hurt yields, increase prices and take up unnecessary foundry production space.

We have already seen the "building blocks" of the M5 series, so it will only make sense to use those same blocks and leverage that at scale across their products.

Apple should just use more 20-core and 40-core GPU tiles (that they already produce). It could be 1 CPU tile and 3 GPU tiles or the classic 2 CPU tiles and 2 GPU tiles.

That's in essence what the 2.5D interposer/packaging technology is about. Leveraging chiplets to increase yields, better thermal control and ultimately lower production costs.

 

[chars1ub0w]

Apple should just use more 20-core and 40-core GPU tiles (that they already produce). It could be 1 CPU tile and 3 GPU tiles or the classic 2 CPU tiles and 2 GPU tiles.

Notice Apple already uses more than one CPU tile variant.
For the base M5, we can assume one is the "binned" version of the other - losing one S-core (defective).
For the M5 pro, the CPU tile is obviously different from the base M5 tile.
We can assume perhaps the same thing happens - one is the binned version - losing one S-core and two P-cores.
(Picture borrowed from Wikipedia.)

Now suppose we use just one CPU tile for the M5 Ultra (to make room for more GPU tiles).
Then it'd be the 6 S-core 12 P-core one, which sounds underpowered.
One solution to keep to one CPU tile is to make a new (more powerful) one.
If I understand it, memory is accessed through the GPUs (controllers are on that tile).
Double the GPU tiles, double the max bandwidth and max amount of RAM.

Actually, the top M5 Max configuration has exactly twice the max memory (128GB) and GPU cores of the M5 Pro.
So Apple already is using two identical GPU tiles here, if that is the GPU building block.
To get 512GB max RAM, you'd need 4 identical GPU tiles for (4 x 20 =) 80 GPU cores.

 

[Pressure]

Notice Apple already uses more than one CPU tile variant.
For the base M5, we can assume one is the "binned" version of the other - losing one S-core (defective).
For the M5 pro, the CPU tile is obviously different from the base M5 tile.
We can assume perhaps the same thing happens - one is the binned version - losing one S-core and two P-cores.
(Picture borrowed from Wikipedia.)
View attachment 2637383
Now suppose we use just one CPU tile for the M5 Ultra (to make room for more GPU tiles).
Then it'd be the 6 S-core 12 P-core one, which sounds underpowered.
One solution to keep to one CPU tile is to make a new (more powerful) one.
If I understand it, memory is accessed through the GPUs (controllers are on that tile).
Double the GPU tiles, double the max bandwidth and max amount of RAM.

Actually, the top M5 Max configuration has exactly twice the max memory (128GB) and GPU cores of the M5 Pro.
So Apple already is using two identical GPU tiles here, if that is the GPU building block.
To get 512GB max RAM, you'd need 4 identical GPU tiles for (4 x 20 =) 80 GPU cores.

It should go without saying that we can disregard the base M5 when talking about the Ultra variant. The base M5 doesn't use chiplets, it's not a heterogenous die. It's a single die on a wafer.

The CPU tile (18-core) is set in stone as are the GPU tiles (20-core and 40-core). The rest is just binning due to yields.

Many users have no use for even more cores as a lot of software isn't really that parallel. GPU cores, however, scale performance. With this new tile approach Apple can offer both solutions without custom dies.

I would like to see Apple offering a variant with only 1 CPU tile and 3 GPU tiles. 18-core CPU and up to 120-Core GPU.

Are you confusing memory PHYs with SLC integrated in the various blocks?

 

[chars1ub0w]

Many users have no use for even more cores as a lot of software isn't really that parallel.

Well, some of us who write code can definitely use way more CPU cores.

Are you confusing memory PHYs with SLC integrated in the various blocks?

eh? I am talking about this.

That is, as some suggest, the M5 Max is using two GPU tiles already, double everything (20 -> 40, 64B -> 126GB). And the Ultra would be x4.

 

[Pressure]

Well, some of us who write code can definitely use way more CPU cores.

eh? I am talking about this.
View attachment 2637391

That is, as some suggest, the M5 Max is using two GPU tiles already, double everything (20 -> 40, 64B -> 126GB). And the Ultra would be x4.

Good point on code compilation 👍🏼 I've seen the 18-core CPU compile WebKit full release in about 15 to 16 minutes.

There is a lot of logic and I/O beside those two tiles, like ANE, Thunderbolt/USB controller, media engine, display engine, ISP, memory PHYs etc etc.

I haven't seen a breakdown of the M5 Pro or M5 Max die, so I can't say for sure but it makes sense (from the press release) that the CPU and GPU are chiplets / tiles or whatever we want to call it.

Can't wait to see a proper die shot of the M5 Pro and M5 Max.

From that x-ray it appears that Apple could be using 20-core GPU tiles (which would make sense) but I have no idea (or if I even identified the correct blocks). The x-ray is very low resolution so hard to guess where logic ends or begins but it could be something like this 🤷🏼‍♂️

But can you see why that rumor you posted doesn't make a lot of sense now?

 

[chars1ub0w]

View attachment 2637396

Yes, that's what I mean.

But can you see why that rumor you posted doesn't make a lot of sense now?

As I said before, I don't know for sure. But Apple has flexibility. It need not use those exact tiles. Why? Because special purpose. For example, M5 Studio tiles might be something they want internally to build out their Private Cloud Compute, and that could go into consumer-ware.