Comparing M1 16GB…

[Warped9]

As I look forward to eventually upgrading from my 2011 21.5 iMac to a new Mac I’ve been trying to get a handle on how the unified RAM, more specifically 16GB, compares to conventional RAM. Yes, I understand the M series are more efficient than Intel or AMD chips, but what is the performance gain, if any, with the unified RAM.

16GB is 16GB, but because it’s unified with the processor does that give it a performance gain or advantage over conventional 16GB or even 32GB RAM?

 

[AdamBuker]

The unified memory architecture of Apple Silicon generally performs better than conventional memory because lower latency and little to no memory swap between CPU and GPU. As far as capacity, it really depends on how you need to use your mac. If all you need is something to browse the web, productivity apps, and/or casual gaming 8-16 GB will be enough. If you want to do light photo/video editing or simple music production or more intense gaming then 16 GB is the way to go. If you need to do professional photo/video work or professional music production or need to use virtual machines, then I would recommend 32 GB minimum. As fast as SSD swap is now, many consumers won’t experience the same slowdown when maxing out their memory usage provided that there’s a decent amount of free space available on their SSD, but in many professional applications you will.

If in doubt, buy the max ram and storage your budget will allow.

 

[Zest28]

"Unified memory" is more efficient as all the data is in 1 place, but you need more RAM in the end as the GPU doesn't have dedicated VRAM.

Btw, if your 2011 iMac uses integrated graphics, it is already using "unified memory".

 

[cupcakes2000]

Btw, if your 2011 iMac uses integrated graphics, it is already using "unified memory".

It’s simply sharing the system memory rather than having onboard memory. Unified memory is a completely different concept.

 

[Spindel]

"Unified memory" is more efficient as all the data is in 1 place, but you need more RAM in the end as the GPU doesn't have dedicated VRAM.

Btw, if your 2011 iMac uses integrated graphics, it is already using "unified memory".

No his 2011 mac has a part of the system memory dedicated to the GPU, it’s walled off and assets that both gpu and cpu need to work on will recide two times in the ram, one copy in the GPU dedicated pool and one copy in the system pool. And they will be copied back and forth when gpu or cpu does changes to their copies.

In the AS macs the assets only exist one time in ram and is accessible by both cpu and gpu for doing changes.

 

[Warped9]

So my suspicion seems to be correct. I’m better off with a machine that has at least the same 32GB RAM as I have now even with the M1 than an M1 with 16GB RAM vs. My i7 with 32GB DDR3 RAM. I was thinking an M1 with 16GB would give me at least the same performance, perhaps somewhat better, than what I have now. But since I cannot user upgrade the M1 RAM it’s better to future proof myself with more memory at the onset.

Thats why I was holding off going for an M1 iMac with 16GB and waiting for the upgrade (likely M3) with (hopefully) 24 or 32GB RAM as an available option. If that doesn’t materialize, even though I like the AIO configuration, then it I’ll most likely go for a base Mac Studio.

 

[cupcakes2000]

So my suspicion seems to be correct. I’m better off with a machine that has at least the same 32GB RAM as I have now even with the M1 than an M1 with 16GB RAM vs. My i7 with 32GB DDR3 RAM. I was thinking an M1 with 16GB would give me at least the same performance, perhaps somewhat better, than what I have now. But since I cannot user upgrade the M1 RAM it’s better to future proof myself with more memory at the onset.

Thats why I was holding off going for an M1 iMac with 16GB and waiting for the upgrade (likely M3) with (hopefully) 24 or 32GB RAM as an available option. If that doesn’t materialize, even though I like the AIO configuration, then it I’ll most likely go for a base Mac Studio.

Well the system offer’s dramatically faster ram in many situations, and potentially dramatically more than one is used to for gpu usage - but it’s not magical ram. If you need 32gb you need 32gb of any ram.

 

[wilberforce]

So my suspicion seems to be correct. I’m better off with a machine that has at least the same 32GB RAM as I have now even with the M1 than an M1 with 16GB RAM vs. My i7 with 32GB DDR3 RAM. I was thinking an M1 with 16GB would give me at least the same performance, perhaps somewhat better, than what I have now. But since I cannot user upgrade the M1 RAM it’s better to future proof myself with more memory at the onset.

Thats why I was holding off going for an M1 iMac with 16GB and waiting for the upgrade (likely M3) with (hopefully) 24 or 32GB RAM as an available option. If that doesn’t materialize, even though I like the AIO configuration, then it I’ll most likely go for a base Mac Studio.

You haven't stated how much memory you typically use, not how much you have. Someone who has 32GB but only typically uses 16GB (not counting cached files), is going to get little performance benefit from upgrading from 16GB to 32GB.

 

[stradify]

Warped9,
What's your memory pressure in Activity Monitor when you're 'pedal to the metal'?

 

[unrigestered]

unified memory is more efficient if you are using your systems with an eye on graphics performance, for stuff that's not utilizing your GPU much, it's benefits are negligibly small at best, if not even a (admittedly small) disadvantage.

 

[cupcakes2000]

unified memory is more efficient if you are using your systems with an eye on graphics performance, for stuff that's not utilizing your GPU much, it's benefits are negligibly small at best, if not even a (admittedly small) disadvantage.

Disadvantage? What do you mean? I don’t see any disadvantage at all, aside it being non user expandable but that’s pretty common for macs regardless.

 

[Warped9]

You haven't stated how much memory you typically use, not how much you have. Someone who has 32GB but only typically uses 16GB (not counting cached files), is going to get little performance benefit from upgrading from 16GB to 32GB.

Warped9,
What's your memory pressure in Activity Monitor when you're 'pedal to the metal'?

I will have to look. I have never actually referred to the Activity Monitor. But I do know with SketchUp as my models get more complex the response to commands, such as zooming in or out on an object, begins to lag. It’s tolerable as it is presently, but it shouldn’t be like that. Rendering is not as fast as it could be. It certainly doesn’t take hours, but it can take several minutes depending on the model.

Now I have looked up reviews of using SketchUp (and other 3D programs) on an M1 iMac. The impression I get is that an M1 with 16GB would essentially match the performance I have now. Maybe a bit better. The same or “a bit better” is not enough of a performance gain for the money over my present 11 year old setup.

I understand SketchUp uses a single processor core, and the higher clock frequency of the processor the better the performance. From what I know the M1 is supposedly about 20% or so faster than the Intel i7 2600S I have in my iMac. I don’t know how much faster an Apple 512GB or 1TB SSD is over my Samsung Evo 500GB SSD. Then, of couse, is the RAM we’re discussing.

 

[stradify]

When you open Activity Monitor click on the Memory tab at the top.
It will show you all your open apps and the amount of RAM they're using.
Being a database it's sortable and searchable.
At the bottom of the page you'll see something like this: Memory Pressure Green is good. No swap used.

 

[unrigestered]

Disadvantage? What do you mean? I don’t see any disadvantage at all, aside it being non user expandable but that’s pretty common for macs regardless.

disadvantages
- memory bandwidth needs to be shared between CPU and GPU (though split memory pools require constant moving of data between them too)
- slightly less RAM available, as some things could have been stored in the dedicated video RAM

it's not an awful lot and probably not noticeable in practice, only measurements might show you a couple of % less on paper

 

[dmccloud]

disadvantages
- memory bandwidth needs to be shared between CPU and GPU (though split memory pools require constant moving of data between them too)
- slightly less RAM available, as some things could have been stored on the dedicated video RAM

it's not an awful lot and probably not noticeable in practice, only measurements might show you a couple of % less on paper

1. The memory bandwidth is actually significantly higher than in most traditional systems, while also eliminating a lot of the latency issues present with off-die RAM like the majority of PCs run. Since this RAM is on the SoC die, it's running at the CPU speed rather than the memory bus speed like traditional systems do. You also forget that in most systems with integrated graphics, the system has to have two copies of data (one for CPU, one for GPU), and then reconcile differences between the two copies. That introduces additional levels of latency and bandwidth congestion that are not present under unified memory (where both the CPU and GPU can access the same data simultaneously).

2. There is actually MORE RAM available to both the GPU and CPU. In a normal AMD or Intel system with 8GB RAM and integrated graphics, around 2GB of RAM is walled off exclusively for the GPU, with the remaining 6GB exclusively reserved for the CPU. In situations where either the CPU or GPU have maxed out their allotment of RAM, there is no means of reallocating RAM from CPU to GPU or vice-versa. With Unified Memory, the system can reallocate as needed, thereby preventing both the CPU and GPU from becoming memory-bound in most cases.

 

[Spindel]

I will have to look. I have never actually referred to the Activity Monitor. But I do know with SketchUp as my models get more complex the response to commands, such as zooming in or out on an object, begins to lag. It’s tolerable as it is presently, but it shouldn’t be like that. Rendering is not as fast as it could be. It certainly doesn’t take hours, but it can take several minutes depending on the model.

Now I have looked up reviews of using SketchUp (and other 3D programs) on an M1 iMac. The impression I get is that an M1 with 16GB would essentially match the performance I have now. Maybe a bit better. The same or “a bit better” is not enough of a performance gain for the money over my present 11 year old setup.

I understand SketchUp uses a single processor core, and the higher clock frequency of the processor the better the performance. From what I know the M1 is supposedly about 20% or so faster than the Intel i7 2600S I have in my iMac. I don’t know how much faster an Apple 512GB or 1TB SSD is over my Samsung Evo 500GB SSD. Then, of couse, is the RAM we’re discussing.

Sounds more like a weak GPU issue than a RAM issue with what you describe with SketchUp. Also if it is an RAM issue the UMA of the AS will help since you don't have to move the data between the GPU pool and the CPU pool since the model data can be accessed by both with UMA.

I personally went from a iMac 27" 2013 with a GTX 780M (4 GB of RAM) and 24 GB of RAM to a M1 Mini 16 GB and under no circumstances has the iMac beaten the M1 Mini in performance (including graphics performance) except for applications that simply don't run on the M1 mini (witch is none for me personally). Even programs under Rosetta runs better on the M1 Mini.

And to be really honest 90% of the time a 8GB M1 Mini would be OKish for me simply because the SSD is fast enough that swaping actually works OK.

 

[Warped9]

So here is what I see. The only app running, other than Finder and Activity Monitor, is SketchUp.

 

[richard371]

16GB vs. 32GB MacBook Pro: How Much RAM is Enough?

Apple's high-end 14-inch and 16-inch MacBook Pro offers the M1 Pro chip with 16GB of RAM as standard, but you can upgrade to 32GB of RAM for...

www.macrumors.com

 

[leman]

So many misconceptions here…

"Unified memory" is more efficient as all the data is in 1 place, but you need more RAM in the end as the GPU doesn't have dedicated VRAM.

This is baseless speculation, and very likely wrong.

It’s simply sharing the system memory rather than having onboard memory. Unified memory is a completely different concept.

It’s the same thing in modern implementations. Intel and AMD had unified memory on their SoCs since 2012 or so.

So my suspicion seems to be correct. I’m better off with a machine that has at least the same 32GB RAM as I have now even with the M1 than an M1 with 16GB RAM vs. My i7 with 32GB DDR3 RAM.

Maybe, maybe not. Just because your old system has more RAM does not actually mean that you need it. Apple offers free returns, why not get a 16GB system and try it out?

 

[unrigestered]

It’s the same thing in modern implementations. Intel and AMD had unified memory on their SoCs since 2012 or so.

depending on who was responsible for the actual design (most likely either intel or MS), unified memory already became a big thing in the first xbox from 2000 (or 2001, i forget). at the very least intel was involved by providing their P3 Celerons (iirc)

edit: forgot about the N64 which already had UMA in 1996, albeit with several bottlenecks inside the whole hardware design

 

[Yebubbleman]

As I look forward to eventually upgrading from my 2011 21.5 iMac to a new Mac I’ve been trying to get a handle on how the unified RAM, more specifically 16GB, compares to conventional RAM. Yes, I understand the M series are more efficient than Intel or AMD chips, but what is the performance gain, if any, with the unified RAM.

16GB is 16GB, but because it’s unified with the processor does that give it a performance gain or advantage over conventional 16GB or even 32GB RAM?

In a traditional system (i.e. any PC or any pre-Apple-Silicon Mac), each component has to access RAM separately and often in their own corner of the RAM. In said traditional systems with discrete graphics processors, said graphics processors have their own RAM. Data needs to travel across the entire system to get from one component to the other. Even in cases where, like on any 2011-2020 Intel 13-inch MacBook Pro, there is no discrete GPU and the RAM is shared, data often has to leave the GPU's portion of RAM, go to the CPU, and then back onto the CPU's portion of RAM. What I just wrote out there is very crude, but I'd imagine you sort of get the basic gist.

With Unified Memory, as Apple has it, the RAM is there on the System-on-a-Chip and all processing/co-processing/sub-processing components can access that data at once without having to move it to and from anywhere.

It's the difference between you having your desk with the papers you manage with your co-workers having the papers they manage on their desk and you all trading these papers around as such...VERSUS all of you working from the same communal desk and having each others papers evenly and perfectly within reach of everyone.

 

[Warped9]

So unless I’m reading my Activity Monitor wrong my system is not overworked. It might possibly be, as suggested upthread, a weak CPU. SketchUp recommends a 4-6 core processor with 3.0 GHz. My i7 is 2.8 GHz. That doesn’t seem like a big difference, but maybe it‘s enough of a difference?

I know when I'm rendering the model the back of the computer (on the left side) gets warm and I can hear the fan although it's not loud.

 

[ian87w]

RAM is RAM. Speed != capacity.
Best way to check, on your current system, run all the apps in your daily workflow, use your computer for a while, then check activity monitor. See where it’s at. If it’s still consuming less than half with no swap, then maybe you can live with 16GB. If not, get the same amount of RAM. You cannot upgrade this later, and future software would only consume more RAM.

 

[theorist9]

The unified memory architecture of Apple Silicon generally performs better than conventional memory because lower latency and little to no memory swap between CPU and GPU.

1. The memory bandwidth is actually significantly higher than in most traditional systems, while also eliminating a lot of the latency issues present with off-die RAM like the majority of PCs run.

I understand avoiding memory swap between CPU and GPU is advantageous, and that AS's memory bandwidth is quite high. But is it really the case that AS's memory latency is lower than that of high-end PC systems? Have you seen reliable measurements that demonstrate this?

Yes, AS RAM is located quite close to the CPU and GPU but, as I'm sure you know, there's more to RAM latency than just proximity.

I found this graph by Andrei Frumusanu at https://www.anandtech.com/show/17024/apple-m1-max-performance-review/2 showing the memory latency of the M1 Max, but I've not been able to find anything equivalent for PC's. According to a previous Anandtech article, the key 'figure of merit' is the "full random" value, which in this case is 111.1 ns:

 

[Zest28]

So many misconceptions here…



This is baseless speculation, and very likely wrong.



It’s the same thing in modern implementations. Intel and AMD had unified memory on their SoCs since 2012 or so.


Maybe, maybe not. Just because your old system has more RAM does not actually mean that you need it. Apple offers free returns, why not get a 16GB system and try it out?

What baseless speculation? I have seen the GPU of my M1 Max pull 21GB RAM. So you need more RAM with M1 chips as the GPU will eat into the RAM. There is no dedicated VRAM.