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

california_kid

macrumors 6502a
Original poster
Sep 9, 2016
544
1,043
San Francisco
Besides the obvious benefit of more physical memory if running multiple apps, I'm wondering if the "Up to 400 GB/s" bandwidth is only applicable to the 64 GB option. From the pics, the M1 Pro only support 2 memory chips for 32 GB max. The M1 Max goes up to 4 chips and 64 GB max. One way to double the bandwidth from M1 Pro would be to double the memory bus width which would require 2 memory chips per bank versus 1, hence 4 total. The other way to double the bandwidth would be to increase the bus speed 2x. If it's just faster bus speed, I personally would save the $400, but if the only way to get 400 GB/s is to upgrade to 64 GB, then I would do that since it would be a major improvement in both graphic and cpu performance.

Anyone know for sure or make an educated guess?
 
Can you please explain what that means? I apologize but I am not a technical guy.
I was concerned that "up to 400 GB/s" meant only for 64 GB configuration but all indications are that either 32 or 64 GB give the full 400 GB/s bandwidth. By asking about the use of 4 chips (Max) or 2 chips (Pro), I was wondering how the Max achieves 2x the bandwidth. So far all the images show 4 memory chips for Max so that could mean that the Max has 2x the number of memory channels versus Pro which only need 2 chips.
 
I was concerned that "up to 400 GB/s" meant only for 64 GB configuration but all indications are that either 32 or 64 GB give the full 400 GB/s bandwidth. By asking about the use of 4 chips (Max) or 2 chips (Pro), I was wondering how the Max achieves 2x the bandwidth. So far all the images show 4 memory chips for Max so that could mean that the Max has 2x the number of memory channels versus Pro which only need 2 chips.
Thanks for the clarification!

However, I still don‘t get the benefit of a 400 GB/s bandwidth over 200. In which scenarios does it matter?
 
Thanks for the clarification!

However, I still don‘t get the benefit of a 400 GB/s bandwidth over 200. In which scenarios does it matter?

Both CPU and GPU use the same unified memory so they have to share bandwidth. If you get 32 GPU cores, that is 4x the number of cores than the original M1 so you want to scale the bandwidth so all the cores can process in parallel. So basically, if you run high-end graphics games and/or apps, the higher bandwidth gives more frames per second. If you look at a high end discrete GPU like an RTX 3080, that GPU uses >700 GB/s bandwidth just for itself!
 
Thanks. Three speculative follow-up questions:

1) I understand that, with more GPU cores, you need more memory bandwidth to maintain the same bandwidth/core. But with a 24-core Max, do the 24 cores benefit from having higher bandwidth/core than in the 32-core Max? I.e., might the per-core GPU performance in the 24-core Max be higher than in the 32-core Max?

2) Since this is unified memory, should the higher bandwidth of the Max also benefit CPU performance (since you now have twice the memory bandwidth per CPU core vs in the Pro)? Or is higher memory bandwith mainly needed on the GPU side?

3) I understand from what you've written that, when Apple said "up to 400 GB/s memory bandwidth" for the Max, the "up to" didn't refer to any dependency on no. of GPU cores or memory size. The qualifier referring to something else. But what else might that be?
 
Thanks. Three speculative follow-up questions:

1) I understand that, with more GPU cores, you need more memory bandwidth to maintain the same bandwidth/core. But with a 24-core Max, do the 24 cores benefit from having higher bandwidth/core than in the 32-core Max? I.e., might the per-core GPU performance in the 24-core Max be higher than in the 32-core Max?

2) Since this is unified memory, should the higher bandwidth of the Max also benefit CPU performance (since you now have twice the memory bandwidth per CPU core vs in the Pro)? Or is higher memory bandwith mainly needed on the GPU side?

3) I understand from what you've written that, when Apple said "up to 400 GB/s memory bandwidth" for the Max, the "up to" didn't refer to any dependency on no. of GPU cores or memory size. The qualifier referring to something else. But what else might that be?

1) In worse case, 32 core performance will be the same as 24, but it will never be less than 24 regardless of memory bandwidth. Imagine a 32 car train versus a 24 car train (cores) running on the same speed track (bandwidth). If each car can carry 10 people (data), then 32 will always carry more since it has more cars. But if only 240 or less people need to get on the train, both the 32 and 24 car train will perform equally since each train can carry all the passengers in one trip.

2) Currently GPU is more memory bandwidth intensive. As the linked article above describes the M1 Max is nearly twice the size of M1 Pro with most of the extra real estate dedicated to the GPU cores.

3) "Up to" could mean several things. One it could mean different memory operations can go up to 400 GB/s but not all. For example reads are general much faster than writes or large amount of small data files take more time to transfer versus one large data file (random versus sequential). It could also mean that the M1 Max typically doesn't need the full speed so to conserve battery life, it may throttle down the speed with no perceived drop in performance.
 
I believe you misunderstood my question #1. When I wrote it was going to include the following as an additional explanatory qualifier, but left it out:

"I understand the 24-core GPU performance can't exceed that of the 32-core model. That's not what I'm asking. I'm asking if the extra per-GPU-core memory bandwidth of the 24-core variant will enable higher per-core GPU performance than in the 32-core model. I assume there is a point at which higher per-core memory bandwidth doesn't help GPU performance for this system—a point beyond which these GPU cores become "saturated" as far as memory bandwidth is concerned. But I have no ideal what that point would be."

Of course, this may be a technical question that can't be answered at this point.
 
I believe you misunderstood my question #1. When I wrote it was going to include the following as an additional explanatory qualifier, but left it out:

"I understand the 24-core GPU performance can't exceed that of the 32-core model. That's not what I'm asking. I'm asking if the extra per-GPU-core memory bandwidth of the 24-core variant will enable higher per-core GPU performance than in the 32-core model. I assume there is a point at which higher per-core memory bandwidth doesn't help GPU performance for this system—a point beyond which these GPU cores become "saturated" as far as memory bandwidth is concerned. But I have no ideal what that point would be."

Of course, this may be a technical question that can't be answered at this point.

Unfortunately no, if there is a scenario where 24 cores and respective per memory bandwidth is optimal, the 32 core GPU will simply shut down 8 cores and only use 24. Just because the CPU and GPU have multiple cores, it doesn't mean all the cores are being used so a 32 core GPU can operate exactly like a 24 core GPU if that is what the system needs. However, vice versa is not true hence the premium price for 32.
 
I was concerned that "up to 400 GB/s" meant only for 64 GB configuration but all indications are that either 32 or 64 GB give the full 400 GB/s bandwidth. By asking about the use of 4 chips (Max) or 2 chips (Pro), I was wondering how the Max achieves 2x the bandwidth. So far all the images show 4 memory chips for Max so that could mean that the Max has 2x the number of memory channels versus Pro which only need 2 chips.
Note that the die shot images aren't showing the memory chips. They're showing the memory controllers. The memory chips themselves are off-die.
 
Two of the workers at Apple explains more about the processors here:
Apple VP — M1 Pro / Max Questions Answered! - YouTube

"Tom Boger, Vice President of iPad & Mac Product Marketing and Tim Millet, Vice President of Platform Architecture, joins to talk about what they thought when Apple first decided to switch the Mac to custom silicon, what it was like bringing their low/slow/wide approach to a thermal envelope as big as the new MacBook Pro, how scalable architecture really scales up this much (and more), how they think about transistor budget in an increasing post-big compute core world, gaming on Mac, and which MBPs we’re all rocking !"
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.