M3-based Mac Studio GPU specs

[Kronsteen]

Hello everybody

I’d appreciate hearing any thoughts you have on the likely spec of the M3-based Mac Studio.

Background: I have a 2013 Mac Pro that I use for a couple of workloads implemented using OpenCL (as well as an Nvidia Jetson Orin developer kit on which I run CUDA-based versions). I am keen to try running the same workload on Apple Silicon, and I’d like to retire my old Mac Pro before security updates for Monterey end, but there is no great urgency, so I will probably hold off until the M3 version of the Studio is available, and will then create a Metal-based version of my code.

Now, the way I portion the workload requires that the number of work items (to use the OpenCL terminology) be a power of two. I was therefore hoping that the base version of the M3 Max would have 32 GPU cores (and that the Ultra would likely then be 64), so I’m a little disappointed that it turns out to have 30.

So my main question is: is there any realistic prospect that the “basic” M3 Max version of the Studio might have 32 GPU cores, or is it inevitable that it will be precisely the same as the M3 Max that has just been announced for the MacBook Pro?

Thanks …. Andrew

 

[Pressure]

Just look at the two top spec M3 Max laptops.

14-core CPU	16-core CPU
30-core GPU	40-core GPU
36GB Unified Memory (3 x 12GB)	48GB Unified Memory (4 x 12GB)

They might cheap out on storage and offer them with just 512GB SSD Storage.

 

[Kronsteen]

Just look at the two top spec M3 Max laptops.

14-core CPU	16-core CPU
30-core GPU	40-core GPU
36GB Unified Memory (3 x 12GB)	48GB Unified Memory (4 x 12GB)

They might cheap out on storage and offer them with just 512GB SSD Storage.

Thanks for the quick reply and, yes, I’d seen the laptop specs. It’s the 30 cores in the ‘base’ M3 Max that I’m a little disappointed by, hence my question about whether there’s any realistic prospect of the M3 Max in the Studio having 32 (or whether it’s inevitable that it will be 30 like the laptops).

(The point being that my workload needs to have a number of work items / threads that is a power of two, so 32 would have been much better. Of course, if I decide to go ahead and buy an M3 Max or Ultra Studio, I could cough up the extra $ for the 40 (or 80) core version, and leave 8 (or 16) of them idle, but it would have been nice not to have had to do that.)

Andrew

 

[ChrisA]

Hello everybody

I’d appreciate hearing any thoughts you have on the likely spec of the M3-based Mac Studio.

Background: I have a 2013 Mac Pro that I use for a couple of workloads implemented using OpenCL (as well as an Nvidia Jetson Orin developer kit on which I run CUDA-based versions). I am keen to try running the same workload on Apple Silicon, and I’d like to retire my old Mac Pro before security updates for Monterey end, but there is no great urgency, so I will probably hold off until the M3 version of the Studio is available, and will then create a Metal-based version of my code.

Now, the way I portion the workload requires that the number of work items (to use the OpenCL terminology) be a power of two. I was therefore hoping that the base version of the M3 Max would have 32 GPU cores (and that the Ultra would likely then be 64), so I’m a little disappointed that it turns out to have 30.

So my main question is: is there any realistic prospect that the “basic” M3 Max version of the Studio might have 32 GPU cores, or is it inevitable that it will be precisely the same as the M3 Max that has just been announced for the MacBook Pro?

Thanks …. Andrew

I assume you want 32 "performance cores" and not just 32 "cores". So to et what you want, you'd need to buy the M3 that Apple advertizes as having 40 or more cores.

Also while yoir app might have 32 threads, it is not the only things running on the Mac, So you need enough performance cores such that your 32 threads never have to wait for a free performance core. You likey want at least a 40-core Mac.

What if you have say 48 cores? Would splitting your task into 64 threads and running on 48 cores be faster than splitting into 32 cores and running on 32 cores? My guess is the 48-core solution would be faster.

But I have a question. Are you writing this in C++ or Swift, or what? and how do you lock a thread to a specified kind of core?

Finally what is the task being done? Could this run on a GPU or Neural core?

 

[fathergll]

So my main question is: is there any realistic prospect that the “basic” M3 Max version of the Studio might have 32 GPU cores,

Anything is possible but there is no realistic prospect of it not matching the MacBook Pro. I wouldn't even waste time thinking about it other than planning on spending the next 7 or so months figuring out how to get the extra $300 for it.

 

[CWallace]

So my main question is: is there any realistic prospect that the “basic” M3 Max version of the Studio might have 32 GPU cores, or is it inevitable that it will be precisely the same as the M3 Max that has just been announced for the MacBook Pro?

For economy of scale and production simplification reasons, I would expect the M3 MAX in the Studio to match the M3 MAX in the MacBook Pro - so the base will offer 30 GPU cores.

 

[Kronsteen]

Anything is possible but there is no realistic prospect of it not matching the MacBook Pro. I wouldn't even waste time thinking about it other than planning on spending the next 7 or so months figuring out how to get the extra $300 for it.

For economy of scale and production simplification reasons, I would expect the M3 MAX in the Studio to match the M3 MAX in the MacBook Pro - so the base will offer 30 GPU cores.

Thanks for the replies. I'm pretty sure you're both correct 👍 (although I wish you weren't .... ). Any prospect of the Studio being different is wishful thinking on my part.

Unfortunately, if I can persuade myself to buy an M3 Ultra Studio, it will quite likely be more than $300 more to get the 80 cores. Here in the UK, the extra 16 GPU cores on the M2 Ultra (to take it from 60 to 76) costs £1,000 (pounds sterling, so about $1,200 USD at current exchange rate) .

Still, I can always dream that the price difference for the extra GPU cores for the M3 will be less ....

 

[galad]

work item != gpu core anyway.

 

[Kronsteen]

I assume you want 32 "performance cores" and not just 32 "cores". So to et what you want, you'd need to buy the M3 that Apple advertizes as having 40 or more cores.

Also while yoir app might have 32 threads, it is not the only things running on the Mac, So you need enough performance cores such that your 32 threads never have to wait for a free performance core. You likey want at least a 40-core Mac.

What if you have say 48 cores? Would splitting your task into 64 threads and running on 48 cores be faster than splitting into 32 cores and running on 32 cores? My guess is the 48-core solution would be faster.

But I have a question. Are you writing this in C++ or Swift, or what? and how do you lock a thread to a specified kind of core?

Finally what is the task being done? Could this run on a GPU or Neural core?

Thanks for taking the time to reply. If I may answer your last two questions first:

My code is already GPU-based (hence my interest in creating a Metal version). So, in this context, I'm interested in GPU cores only.

The two programs that I would like to try first are (a) some linear algebra working on what might loosely be described as binary matrices; and (b) number-theoretic transforms (essentially a modular arithmetic version of FFTs) using long vectors of integers. They are written in C++ and OpenCL / CUDA (OpenCL for the version that runs on my old 2013 Mac Pro and Intel MacBook Pro, CUDA for my Nvidia Jetson developer kit).

So, yes, I could partition the work the way you describe but the kernel(s) would have to run sequentially in order to run the 64 units of work (even though the unit of work would be half the size vs. a 32-way split), given the SIMD model of execution on the GPU, so I don't think it would be any quicker. Indeed, given the possibility of increased memory contention, it might even be slower, although it's also possible that, if the GPU was being clever, it could execute one batch of threads while the other was waiting for data to materialise. So perhaps worth trying.

You do make a good point about other things running although, for this entirely non-visual stuff that I play with, there is precisely nothing happening on the display while the code is running. (Actually, on my laptop, the code runs on the dGPU, and I imagine that it may be the integrated Intel GPU that is handling mundane Terminal interactions on the screen?). Nonetheless, there's never any harm in having a little spare capacity .... .

Andrew

 

[Kronsteen]

work item != gpu core anyway.

Yes, thanks, I realise that. But, of course, the number of execution units (is that the right word in Metal terminology?) in each GPU core is, I believe, a power of two. Strictly speaking, it's the overall number of execution units that I'm interested in, but the number of GPU cores is obviously a factor of that.

 

[diamond.g]

Yes, thanks, I realise that. But, of course, the number of execution units (is that the right word in Metal terminology?) in each GPU core is, I believe, a power of two. Strictly speaking, it's the overall number of execution units that I'm interested in, but the number of GPU cores is obviously a factor of that.

1024 ALU's per Core
EDIT: at least for M1 and M2 we assume the GPU in M3 has the same ALU per core count).

 

[Fomalhaut]

Anything is possible but there is no realistic prospect of it not matching the MacBook Pro. I wouldn't even waste time thinking about it other than planning on spending the next 7 or so months figuring out how to get the extra $300 for it.

It's an extra $500, because you are obliged to get a minimum of 48GB RAM when you choose the 16/40 core M3 Max.

 

[Kronsteen]

It's an extra $500, because you are obliged to get a minimum of 48GB RAM when you choose the 16/40 core M3 Max.

Yes, good point, I guess that explains it being $500.

What’s irritating me about the Studio pricing — and may well end up with my deciding to forgo the Ultra Studio and buy an M3 Max MacBook Pro instead — is that, here in the UK, Apple charge an extra £1,000 (UKP, so ~$1,200) to go from the 60 to 76 core M2 Ultra, even though going from 30 to 38 only costs £200 more.

I may be missing something, but that just doesn’t feel reasonable to me. I don’t mind paying for Apple kit that I will get good use out of and will enjoy using (that certainly applied to my 2008 and 2013 Mac Pros and 2012 retina MacBook Pro, although my 2019 16” MacBook Pro has been slightly disappointing), but this sort of pricing disparity just doesn’t make sense to me.

 

[fathergll]

It's an extra $500, because you are obliged to get a minimum of 48GB RAM when you choose the 16/40 core M3 Max.

Yes that is true I just looked over at the chip price but forgot Apple has the ram increase required though probably won't matter if the OP was planning on increasing the memory anyways(I assumed OP would increase based on history of keeping previous Mac for 10 years and going with minimum RAM on a non upgradable machine would be unwise if OP was planning on keeping it for many years).

 

[fathergll]

Yes, good point, I guess that explains it being $500.

What’s irritating me about the Studio pricing — and may well end up with my deciding to forgo the Ultra Studio and buy an M3 Max MacBook Pro instead — is that, here in the UK, Apple charge an extra £1,000 (UKP, so ~$1,200) to go from the 60 to 76 core M2 Ultra, even though going from 30 to 38 only costs £200 more.

I may be missing something, but that just doesn’t feel reasonable to me. I don’t mind paying for Apple kit that I will get good use out of and will enjoy using (that certainly applied to my 2008 and 2013 Mac Pros and 2012 retina MacBook Pro, although my 2019 16” MacBook Pro has been slightly disappointing), but this sort of pricing disparity just doesn’t make sense to me.

Why would you not look to buy a M3 Mac Studio instead of a M3 Max MacBook Pro?

 

[Kronsteen]

Why would you not look to buy a M3 Mac Studio instead of a M3 Max MacBook Pro?

Fair question.

My original intention, before Monday's M3 announcement, had been to wait for an M3 Ultra Studio. At some point, I would most likely also have replaced my 2019 16" MacBook Pro with an M3 (or 4) laptop, too (either base M3 or M3 Pro).

I had been hoping that the lower tier Ultra would be at least 64 GPU cores, so was disappointed to see that it will, very likely, be 60. What then irritates me about the Studio pricing is that currently, here in the UK, going from the lower to upper tier Ultra Studio costs and extra £1,000 (UKP).

So, if I baulk at paying that extra £1,000 (assuming that the M3 Studio pricing is no more favourable), I could, of course, choose the 60 GPU core M3 Ultra Studio, when available.

But it then occurred to me that, instead of buying a (probably 60 rather than 80 GPU core) M3 Ultra Studio and a M3 (or 4), possibly Pro, 16" MBP, I could instead buy only a 40 GPU core M3 Max MBP.

For sure, gets me rather less GPU capacity, but 40 M3 GPU cores will be 4-5 times more powerful (at least on paper) than my current machines. I would have the convenience of just one system -- and would save quite a bit of money, probably £3,000 at least (an M3 Max 16" MBP vs. an M3 Ultra Studio + an M3 (Pro?) 16" MBP).

So, so long as I can satisfy myself that it's ok to run my workloads on a MacBook Pro -- and I'm increasingly confident that it will be -- this begins to feel like a pretty easy decision.

 

[fathergll]

Fair question.

My original intention, before Monday's M3 announcement, had been to wait for an M3 Ultra Studio. At some point, I would most likely also have replaced my 2019 16" MacBook Pro with an M3 (or 4) laptop, too (either base M3 or M3 Pro).

I had been hoping that the lower tier Ultra would be at least 64 GPU cores, so was disappointed to see that it will, very likely, be 60. What then irritates me about the Studio pricing is that currently, here in the UK, going from the lower to upper tier Ultra Studio costs and extra £1,000 (UKP).

So, if I baulk at paying that extra £1,000 (assuming that the M3 Studio pricing is no more favourable), I could, of course, choose the 60 GPU core M3 Ultra Studio, when available.

But it then occurred to me that, instead of buying a (probably 60 rather than 80 GPU core) M3 Ultra Studio and a M3 (or 4), possibly Pro, 16" MBP, I could instead buy only a 40 GPU core M3 Max MBP.

For sure, gets me rather less GPU capacity, but 40 M3 GPU cores will be 4-5 times more powerful (at least on paper) than my current machines. I would have the convenience of just one system -- and would save quite a bit of money, probably £3,000 at least (an M3 Max 16" MBP vs. an M3 Ultra Studio + an M3 (Pro?) 16" MBP).

So, so long as I can satisfy myself that it's ok to run my workloads on a MacBook Pro -- and I'm increasingly confident that it will be -- this begins to feel like a pretty easy decision.

Oh in that case this is a very easy decision with just getting the M3 Max MBP or waiting until next year to see if M4 MBP rumors happen and use the Pro until the very end of OS support.