Is RAM usage equivalent between silicon and Intel macbooks?

[Badger.with.hat]

Do the new silicon Macbook Pros have RAM usage and performance rougly equivalent to their older Intel machines?

If I were to copy over my same setup and workflow to an MX Max macbook, would I see it use less RAM across the board, or would the same basic resource demands extract an identical draw it as they did on the older Intel machines?

I've been hoping to replace my stalwart Mac Pro with a laptop, and I think the performance is finally to the point where I can hook it up to the panoply of monitors, drives, cameras, and attachments I use. For years, I've seen it routinely use 50+ gb of its 64 gb of RAM, to the point where it stood out whenever it would use anything less than 30 gb in daily use. I think at a baseline, upon boot, with nothing running, it uses about 12 to 16gb of RAM just to display from 3 4K+ monitors (1 scaled and 2 vertically oriented, seems to exact a greater performance toll). The fact a 2013 machine could do so is pretty impressive, I guess, I've been spoiled. I know I should be more sharp and tidy in my workflow but.....I'm not gonna, so I want to buy a machine that's very slop-tolerant of my bad habits. I can often see that the WindowServer process is pretty greedy, with just a handful of folders open it's eating 20-50% of my CPU, too.

Sorry if this has been retread, I've dug in with all the keywords I can find and only see people talking about how much better the M-series feels, but not many mentions of people who have run the same setup and workflow and seen their new M1/2/3 only used a third/half/etc of the ram it used to on an intel machine. I'm shopping for a MAX 16" with 64 gb of ram anyways, whenever I get it I'll update this to see how it goes.

Thanks for your time!

 

[Mike Boreham]

MacOS will always use available RAM whether it is 8GB or 128GB. What matters is the Memory Pressure, yours is green at the lowest level the window can show. There is no memory issue. Some discussion and Apple link here.

 

[Nermal]

With unified memory, graphics-related workloads have different memory usage. However, "normal" (compute) workloads would use just as much RAM as they do on Intel, if using the same software.

 

[FreakinEurekan]

Apple Silicon will not "use less RAM" than Intel. It might allocate RMA in different ways but if you needed 64GB on Intel, you're gonna want 64GB with Apple Silicon

 

[Badger.with.hat]

Do the new silicon Macbook Pros have RAM usage and performance rougly equivalent to their older Intel machines?

If I were to copy over my same setup and workflow to an MX Max macbook, would I see it use less RAM across the board, or would the same basic resource demands extract an identical draw it as they did on the older Intel machines?

I've been hoping to replace my stalwart Mac Pro with a laptop, and I think the performance is finally to the point where I can hook it up to the panoply of monitors, drives, cameras, and attachments I use. For years, I've seen it routinely use 50+ gb of its 64 gb of RAM, to the point where it stood out whenever it would use anything less than 30 gb in daily use. I think at a baseline, upon boot, with nothing running, it uses about 12 to 16gb of RAM just to display from 3 4K+ monitors (1 scaled and 2 vertically oriented, seems to exact a greater performance toll). The fact a 2013 machine could do so is pretty impressive, I guess, I've been spoiled. I know I should be more sharp and tidy in my workflow but.....I'm not gonna, so I want to buy a machine that's very slop-tolerant of my bad habits. I can often see that the WindowServer process is pretty greedy, with just a handful of folders open it's eating 20-50% of my CPU, too.
View attachment 2372546

Sorry if this has been retread, I've dug in with all the keywords I can find and only see people talking about how much better the M-series feels, but not many mentions of people who have run the same setup and workflow and seen their new M1/2/3 only used a third/half/etc of the ram it used to on an intel machine. I'm shopping for a MAX 16" with 64 gb of ram anyways, whenever I get it I'll update this to see how it goes.

Thanks for your time!

This is more or less what I'm seeing after the switch and a few weeks to keep it going. The M1 Max does a great job replacing that old Mac Pro can, I'll finally say the long awated day when a laptop could be a workstation without compromise appears to be here!

I seem to see this kind of usage in comparison, not sure which app is so much more economical, but it's good to see. Now I just have to stop worrying about having so much in my data written column....hmm.

 

[Azedquery]

Apple Silicon will not "use less RAM" than Intel. It might allocate RMA in different ways but if you needed 64GB on Intel, you're gonna want 64GB with Apple Silicon

You're 100% correct (which I found out the hard way). And if you had a separate graphics card, the RAM on that needs to be added to the equation on an M-anything. On really huge files my 64GB M3 swaps to SSD when my old 64GB intel + 8GB GPU did not, (tested using exactly the same files).

 

[mr_roboto]

You're 100% correct (which I found out the hard way). And if you had a separate graphics card, the RAM on that needs to be added to the equation on an M-anything. On really huge files my 64GB M3 swaps to SSD when my old 64GB intel + 8GB GPU did not, (tested using exactly the same files).

That's not really how discrete GPU memory works. It doesn't extend the total memory pool available to general purpose programs. It should be viewed as a kind of cache which stores information that's usually also stored in main memory; the graphics card needs this cache because it has to have much faster access to that data than a PCIe link can provide. For example, x16 PCIe gen 3 optimistically provides about 16 GB/s, gen 4 about 32, gen 5 about 64, but Apple's W6900X MPX module for 2019 Intel Mac Pros has 512 GB/s local memory. So all the data the card needs to crunch on gets uploaded to its local memory, but that data started out in main system memory, and often stays there.

I'm not saying you didn't observe what you did, but I suspect there were other factors involved, especially if the swapping was severe.

 

[Azedquery]

That's not really how discrete GPU memory works. It doesn't extend the total memory pool available to general purpose programs. It should be viewed as a kind of cache which stores information that's usually also stored in main memory; the graphics card needs this cache because it has to have much faster access to that data than a PCIe link can provide. For example, x16 PCIe gen 3 optimistically provides about 16 GB/s, gen 4 about 32, gen 5 about 64, but Apple's W6900X MPX module for 2019 Intel Mac Pros has 512 GB/s local memory. So all the data the card needs to crunch on gets uploaded to its local memory, but that data started out in main system memory, and often stays there.

I'm not saying you didn't observe what you did, but I suspect there were other factors involved, especially if the swapping was severe.

Yes, I'm aware of that, but I learned from previous experience that it does affect graphics display. The swapping was very severe on M3 48GB with yellow/red mem pressure. It's less severe on M3 Max 64GB, but still present. This applies specifically to memory-intensive graphics processes such as enlarging files, displaying pixel-edits on large files in real time (without which tools such as drawing, cloning and graduated filters are unusable) and displaying a thumbnail (rather than a white block) in apps like Adobe LR and InDesign.

Both internal Intel and M3M SSDs = 1TB (wish I could have upgraded the M3M to 2TB, but long story). I have yet to install or use all of my usual greedy graphics apps (or work with a bazillion open tabs, as is my wont), so the M3M 64GB should be beating the old Intel hands down, but it isn't.

I'm not a technical bod and I won't pretend to understand the numbers you quoted above, but with same files, same test as before, what other factors do you think might be at play?

 

[thebart]

Mx machines have to use RAM that is usually offloaded to video RAM. I believe depending on your monitor and what resolution you run it at, MacOS may use more RAM to perform scaling. If you look at system monitor memory usage, you see a lot of processes with (GPU) at the end taking quite a bit of memory

This is why I think my 16gb m1 goes into swap and pressured memory while my 12gb PC with dGPU hardly ever does even though I am doing the same work and using the same cross platform apps or equivalent Mac apps. Mac system monitor annoyingly doesn't group together all the processes of an app, so it's hard to tell how much memory the app is using without manually adding up all the individual threads like an animal, but a casual inspection shows that chrome, gitkraken, and sublime text use more memory in Mac than in Windows. Now maybe that's because those apps aren't as optimized for Mac as for Windows (they're all silicon apps, btw), but regardless that's the reality. And Mac native apps use lots of memory too. Warp, a Mac exclusive terminal emulator, uses like almost a gig, probably because it has GPU accelerated graphics.

So from my experience, MacOS needs at least as much if not more memory than Windows. And of course you can't upgrade memory so you better think hard before buying.

 

[mr_roboto]

I'm not a technical bod and I won't pretend to understand the numbers you quoted above, but with same files, same test as before, what other factors do you think might be at play?

Well, to unpack a little, GB/s is gigabytes per second - how fast can data move between two locations? That high-end AMD GPU for 2019 Mac Pros I mentioned can access data in its local memory at 512 GB/s, but if the GPU has to read or write anything in main memory, it goes through a 16 GB/s PCIe connection - 32x slower.

Point being, that PCIe communications bottleneck means whatever memory is on the GPU side isn't terribly useful to the CPU, and vice versa. And since the GPU typically has significantly less memory than the CPU, often whatever's currently stored in GPU memory is just a copy of something in CPU memory, so that the system can upload different data to the GPU as-needed. (Sometimes things can be moved so that they're exclusively stored in GPU memory, of course, and that's where you'd get the effect of extending the total memory capacity of the system.)

What are the sizes of these files? Not necessarily the on-disk size, but the pixel dimensions and color depth?

 

[GuruZac]

How do I find memory pressure? Just curious to see how my workflow utilizes the 18GB of RAM on my MBP.

Edit: Figured out it's in Activity Monitor, which I completely forgot I knew.

 

[XboxEvolved]

I think both will always try to use as much memory as they can, I think with Apple silicon however, due to the ram being integrated, speeds, etc, less ram in many cases can do more for you on a silicon Mac than one with lots of ram because of those reasons.

 

[mr_roboto]

I think both will always try to use as much memory as they can, I think with Apple silicon however, due to the ram being integrated, speeds, etc, less ram in many cases can do more for you on a silicon Mac than one with lots of ram because of those reasons.

No, this isn't true. How much RAM you need to avoid slowdowns is almost entirely a function of the apps you run, and often (but not always) the size of the datafiles you're loading and manipulating with those apps.

Apple Silicon style RAM integration enhances performance and reduces power. If you don't buy enough RAM to avoid swapping in your workloads, your computer will often be sitting there doing no computation, just waiting for data to finish swapping between RAM and disk so it can continue to do real work. No matter how fast your RAM is, you still need enough of it.

 

[Zest28]

A M1/2/3 laptop will use way more RAM because the RAM is shared with the GPU. With intel or AMD, your GPU has VRAM on the GPU itself.

So think again about your 8GB RAM on your Mac, it's not really 8GB that you have available with the GPU also eating into this same pool.

 

[leman]

A M1/2/3 laptop will use way more RAM because the RAM is shared with the GPU. With intel or AMD, your GPU has VRAM on the GPU itself.

So think again about your 8GB RAM on your Mac, it's not really 8GB that you have available with the GPU also eating into this same pool.

This has been discussed many times. GPU memory does not work this way. Since system RAM is used as staging area for the GPU, the driver will maintain a sizeable GPU-related memory allocation on the CPU side as well.

 

[dmccloud]

A M1/2/3 laptop will use way more RAM because the RAM is shared with the GPU. With intel or AMD, your GPU has VRAM on the GPU itself.

So think again about your 8GB RAM on your Mac, it's not really 8GB that you have available with the GPU also eating into this same pool.

Both Intel and AMD integrated GPUs use system RAM for the GPU memory allocation. This is why systems with 8GB often report as 6GB available under Windows (or 16Gb systems report as having 12GB available).

 

[Zest28]

Both Intel and AMD integrated GPUs use system RAM for the GPU memory allocation. This is why systems with 8GB often report as 6GB available under Windows (or 16Gb systems report as having 12GB available).

Not really. If you play video games, you will see that some graphics settings are limited by the VRAM on your GPU and not the amount of RAM you have.

The only time I have seen lower RAM being reported is with integrated GPU's, but I am talking about dedicated GPU's like NVIDIA GPU's obviously.

 

[pshufd]

Not if you're running x86 programs under Rosetta 2.

 

[darngooddesign]

I think you’ll be fine with 64GB. Try it out and if not the return period is around two weeks.

 

[GuilleA]

RAM is RAM. The more, the better. Unified Memory, while being more efficient, is still the same.

 

[Schnort]

Point being, that PCIe communications bottleneck means whatever memory is on the GPU side isn't terribly useful to the CPU, and vice versa.

This is true.

And since the GPU typically has significantly less memory than the CPU, often whatever's currently stored in GPU memory is just a copy of something in CPU memory, so that the system can upload different data to the GPU as-needed.

Also true, mostly.

What's not true is frame/render buffers must be stored somewhere, and you'd prefer it to be separate from your main CPU. There is a ton of traffic that happens with these. (updating them with a new render, and streaming from them to output to the monitor)

In a discrete GPU, they're stored on the GPU local RAM, and usually rendered to from locally held textures (which would be copies of what's in system RAM.

That means all that traffic stays off the main CPU bus and on the GPU.

On the apple unified memory scheme, while there's more CPU bus bandwidth it also is impacted by normal GPU operation. More monitors, higher res, higher refresh = more drain on system memory and available bandwidth.

 

[mr_roboto]

What's not true is frame/render buffers must be stored somewhere, and you'd prefer it to be separate from your main CPU. There is a ton of traffic that happens with these. (updating them with a new render, and streaming from them to output to the monitor)

In a discrete GPU, they're stored on the GPU local RAM, and usually rendered to from locally held textures (which would be copies of what's in system RAM.

That means all that traffic stays off the main CPU bus and on the GPU.

On the apple unified memory scheme, while there's more CPU bus bandwidth it also is impacted by normal GPU operation. More monitors, higher res, higher refresh = more drain on system memory and available bandwidth.

None of that is counter to the point I was making, which was that discrete GPUs don't extend the quantity of memory available in the way some people assume.

But as far as your point about memory quality goes - CPUs typically need a lot less memory bandwidth than GPUs, Apple Silicon GPUs are considerably less bandwidth-hungry than NVidia and AMD discrete GPUs, and Apple does equip all these SoCs with GPU-scale memory performance. The result is that Apple Silicon unified memory works quite well in practice. The fears you're trying to raise don't really happen; people don't experience CPU slowdowns just because they connected more monitors.

 

[leman]

What's not true is frame/render buffers must be stored somewhere, and you'd prefer it to be separate from your main CPU. There is a ton of traffic that happens with these. (updating them with a new render, and streaming from them to output to the monitor)

In a discrete GPU, they're stored on the GPU local RAM, and usually rendered to from locally held textures (which would be copies of what's in system RAM.

That means all that traffic stays off the main CPU bus and on the GPU.

On the apple unified memory scheme, while there's more CPU bus bandwidth it also is impacted by normal GPU operation. More monitors, higher res, higher refresh = more drain on system memory and available bandwidth.

Framebuffer bandwidth impact is pretty much negligible. A full backing buffer frame for the 16” MBP takes 30MB, rendering and displaying that at full 120FPS needs 2x3.5GB/s = 7GB/s (once to stream the FB from the GPU to RAM and once to stream it from RAM to the display controller). That is just a small fraction of the available bandwidth, and will be further reduced by techniques such as framebuffer compression. Besides, for normal desktop composition only dirty screen areas need to be redrawn, further reducing the needed bandwidth. It might have been a problem for base M1, not so much for all other chips.

Something worth noting is that FB transfers on Apple are highly optimized. The GPU can transfer the entire tile at once, which is best-case scenario for the memory controller. And Apples memory QoS system makes sure that the traffic is appropriately prioritized. In fact, I wouldn’t be surprised if the framebuffer never leaves the SOC cache.

Finally, it is not that different in the dGPU laptop world. Many gaming laptops stream the framebuffer to the system RAM and do display output via iGPU to support quick GPU switching.

 

[jonh229]

I came here looking to see what the deal is with unified RAM. My 2017 i7 has 40GB RAM.
Everyone is chatting up how great the new M4 mini is with 16GB base ram. I get how the cpu/gpu's can make it all go fast but I still can't get over how little ram it is. And it's an added $200 to go to 24GB and yet another $200 to max it out at 32GB. But my i7 is still more than that. Pardon the pun, but this just doesn't compute for me. What am I missing?

 

[mr_roboto]

I came here looking to see what the deal is with unified RAM. My 2017 i7 has 40GB RAM.
Everyone is chatting up how great the new M4 mini is with 16GB base ram. I get how the cpu/gpu's can make it all go fast but I still can't get over how little ram it is. And it's an added $200 to go to 24GB and yet another $200 to max it out at 32GB. But my i7 is still more than that. Pardon the pun, but this just doesn't compute for me. What am I missing?

Nothing. Ignore anyone who tells you Apple Silicon Macs need less RAM, it's a dumb myth that was spread early on and has been annoyingly persistent.

That said, there's also a lot of people who blindly followed advice to get lots of RAM, and often they didn't actually need it. So, the key question is: do you know for sure you need 40GB?