32GB M1 RAM can equal 32GB of RAM + 28GB VRAM

[rezwits]

Why do you keep repeating the top card they have only has 12 GB of VRAM that’s dead wrong just stop.

IDK tech changes maybe there is a NEWER one with MORE, IDK, just an example...

Enlighten me, what's the TOP OF LINE GAMING CARD? and how much VRAM does it have??

But why does it even matter how much VRAM it has? /s

But I will confess I was misleading and hyping things up when I said M1 RAM = 2x RAM

I should have titled this "32GB M1 RAM can equal 32GB of RAM + 28GB VRAM"

I just couldn't FIT IT IN THE title , I am just excited ok, damn...LOL

 

[LeeW]

Here, I'll try a different route.

Only if it takes you back out the door you came in.

 

[Eric40962005]

IDK tech changes maybe there is a NEWER one with MORE, IDK, just an example...

Enlighten me, what's the TOP OF LINE GAMING CARD? and how VRAM does it have??

Titan isn’t a gaming card for one and the one you keep referencing is over 3 years old. 24 gigs

 

[rezwits]

Titan isn’t a gaming card for one and the one you keep referencing is over 3 years old. 24 gigs

Wow nice, so there is a gaming card that can take advantage of a PC with 32 GB of RAM! cool
So, 32 GB of RAM and 24 GB of VRAM... cool cool, nice nice

I don't care to keep up with the latest GPUs on the PC side because it's all specs and I am not a gamer, or PC user, plus the "tech" changes and I have other things I am working on... got tired of the "newer" "newer" "newer" every 3 months 6 months etc gpu rat race, and unless I am building a Gaming Rig IRDC... if and when I do (if ever), I'll do 3 months research, then build the RIG. ??

So I stopped caring...

n.b. the examples have nothing to do with my point they were just there for reference... I am sure there is what 3 or 4 different versions of the GPU w24GB? IDK/C, but thanks for the heads up! But I mainly REF'd the 32GB Workstation GPU because of the "Application Here"

 

[project_2501]

comment deleted as others have answered better

 

[Xiao_Xi]

Unified memory is more versatile than traditional split memory because you can distribute the memory between CPU and GPU depending on your workload. You have more freedom to use it, but it doesn't mean you have more.

 

[rezwits]

Just for comedy:

Imagine you have Pizza with 8 slices.

On an Intel in order for the CPU and GPU to "get full" they would have to TAKE and EAT 4 slices each.

On an Apple (Silicon) all the CPU and GPU would have to do to "get full" is LOOK at the PIZZA...
(or the GPU watch the CPU eat it... i.e. just by watching the CPU eat the pizza the GPU is FULL)

LOL haha

n.b. a simple function would be (for certain applications) the amount of GPU RAM you have equals CPU Unified RAM - 4GB
therefore if you have 256 GB of Unified RAM you theoretically at the SAME time have 252 GB of VRAM.
but with NON-graphical applications like loading up an insane EXCEL spreadsheet you don't get this 252 GB, it's non-gpu

In the future, the discussion of how much VRAM you have won't even matter, except on PCs

 

[science_]

Why is such a simple concept being railroaded in here? Stop confusing people.

If your workflow before used 16GB of traditional RAM on intel, then on an M1 you would need to purchase the 32GB model or else you would have no overhead for the GPU to utilize.

If your workflow pushes the CPU and GPU hard then you would benefit from buying as much unified memory as you could. If your workflow is CPU heavy but doesn't require much GPU resources then you do not need to take into account all of this (to an extent of course).

However, I do think everyone should be adding at least 4GB of RAM to their workflow RAM usage when picking how much unified memory to go with. I believe 4GB as a baseline allowed to the GPU would be realistic for casual users.

Of course since it is unified, you also get the benefit of utilizing that entire pool of memory by both GPU and CPU depending on what your workflow is that day.

Take that and extrapolate it for other use cases.

 

[arvinsim]

JFC, another one of those "8GB M1 RAM = 16GB Intel RAM" posts ?‍♂️

 

[Shirasaki]

Ok, so everybody else reading your thread is confused on what you try to say.
Let me clear a bit for you.
Supposedly I have a M1 Max system with 64GB of MEMORY (that can either be RAM or VRAM).
macOS occupies 4GB of MEMORY, so 60GB left for the user.
I open a BIG 3D project that needs 30GB of VRAM. Because M1 Max 64GB is Unified Memory, macOS assigns 30GB of MEMORY as VRAM to the program rendering that 3D project. Now I have 30GB of MEMORY to work with.

I need to open 30 tabs on Safari, 30 tabs on Microsoft Edge and 30 tabs on Google Chrome, which takes 12GB of MEMORY. Since those browsers don't render anything atm, they are considered as RAM. I have 18GB to work with now.

Suddenly I want to open a bunch of high res photos in photoshop for my work, also a few other video projects. Those combined used the remaining 12GB of MEMORY. macOS will give it to the program. Heck, swapping might already happening in the background.

Same task on a PC with a GPU that has 32GB of VRAM and 64GB of RAM. Now, instead of using 30GB of RAM, that 3D project rendering happens solely on GPU, using up all that 32GB of VRAM. However, the CPU cannot directly access those 32GB of VRAM, so some data will need to be transferred to the RAM for processing. This could eat a certain amount of RAM but probably not as much. The rest task happens mostly on RAM instead of VRAM.

Mind you that in the PC world they are introducing a new technology so CPU can access VRAM directly, bypassing RAM. If your logic holds any water, does that mean somehow my 64GB RAM + 32GB VRAM system now has 96GB of... RAM? VRAM? That doesn't make any sense.

What M1 does is that any application can access their data regardless of type, all with high bandwidth. So another program wanting to access the 3D rendering data, they don't have to go all the way out to VRAM, access it, and go all the way back to CPU to process it like what most PC we have nowadays does, improving performance. I suggest you to look up something called resource pool in case you still hold your old thinking.

 

[rezwits]

Ok, so everybody else reading your thread is confused on what you try to say.
Let me clear a bit for you.
Supposedly I have a M1 Max system with 64GB of MEMORY (that can either be RAM or VRAM).
macOS occupies 4GB of MEMORY, so 60GB left for the user.
I open a BIG 3D project that needs 30GB of VRAM. Because M1 Max 64GB is Unified Memory, macOS assigns 30GB of MEMORY as VRAM to the program rendering that 3D project. Now I have 30GB of MEMORY to work with.

...

Mind you that in the PC world they are introducing a new technology so CPU can access VRAM directly, bypassing RAM. If your logic holds any water, does that mean somehow my 64GB RAM + 32GB VRAM system now has 96GB of... RAM? VRAM? That doesn't make any sense.

100% agree with you and I want to add a little more

Let's say with your M1 @ 64GB of MEMORY

You wanted to open a HUGE 3D project that needs 50GB of VRAM, YOU CAN!

And while it's in GPU (space), the CPU can fiddle with it just the same, doesn't have to make a "copy" just to edit "points"...

i.e. in the [traditional] sense or ways, you had to have a "copy" in [RAM], and a copy in [VRAM], the GPUs.

to your 2nd point, yeah I am just hearing about this "new technology" type of CPU directly accessing VRAM, bypassing RAM.
You would be limited still to how much VRAM the GPU has, BUT you wouldn't have to SWAP/HOLD freeing up your MAIN RAM
...which would be KEY!

but your GPU still has to have THIS RAM. The "tech" would be similar, but IN THIS APPLICATION, you're upgrading Side-By-Side.

32+32, where as on the Mac you're just going 64 GB SoC... waaaaaay more versatile! Because sometimes you could open a 50 GB "Data File" and put that in Memory, where with a PC with 32 + 32, you couldn't. Unless your App somehow was able to utilize that 32GB on said GPU (in a non-graphical way)... which mining rigs DO, that is put the BLOCKCHAIN in VRAM! (but for f-s sake let's not talk about mining rigs) LOL

But in either case you would be limited to a 32GB FILE (data set), because I am DAMN sure there is no POOLING of RAM+VRAM of 1 FILE "draped across", other than OUR BOY the M1, (and prior A series chips, but they only had like 1GB, 2GB, 3GB, and 4GB)

 

[cmaier]

100% agree with you and I want to add a little more

Let's say with your M1 @ 64GB of MEMORY

You wanted to open a HUGE 3D project that needs 50GB of VRAM, YOU CAN!

And while it's in GPU (space), the CPU can fiddle with it just the same, doesn't have to make a "copy" just to edit "points"...

i.e. in the [traditional] sense or ways, you had to have a "copy" in [RAM], and a copy in [VRAM], the GPUs.

to your 2nd point, yeah I am just hearing about this "new technology" type of CPU directly accessing VRAM, bypassing RAM.
You would be limited still to how much VRAM the GPU has, BUT you wouldn't have to SWAP/HOLD freeing up your MAIN RAM
...which would be KEY!

but your GPU still has to have THIS RAM. The "tech" would be similar, but IN THIS APPLICATION, your upgrading Side-By-Side.

32+32, where as on the Mac you're just going 64 SoC...

No, you didn’t have to have a copy of whatever is in VRAM in RAM. You are misleading people.

 

[Eric40962005]

No, you didn’t have to have a copy of whatever is in VRAM in RAM. You are misleading people.

Oh that’s what he thinks, makes more sense now when he was talking about 12 GB of VRAM in a pascal titan card, with 16gb of system ram. He just has no idea how it works in general.

 

[cmaier]

Oh that’s what he thinks, makes more sense now when he was talking about 12 GB of VRAM in a pascal titan card, with 16gb of system ram. He just has no idea how it works in general.

Yep. He thinks you mirror the entirety of VRAM in RAM or something.

 

[Eric40962005]

Yep. He thinks you mirror the entirety of VRAM in RAM or something.

Yea I was confused why he kept talking about that titan card no one uses anymore. Yea I wonder where he came up with these ideas.

 

[rezwits]

Yep. He thinks you mirror the entirety of VRAM in RAM or something.

Yes in MY APP, I loaded from SSD -> RAM, then when I "render/rip/shred" it got transferred to the GPU. (ON INTEL, this was the case)

i.e. My APP:

SSD -> RAM (edit/tinker/whatever) -> GPU [Intel chip]
now
SSD -> RAM (edit/tinker/whatever) [Apple chip] (no swap/update needed)

With the Apple it's all the same...

And yes in My App on Intel what was in MEMORY was ASSEMBLED on the GPU.
but then updated in RAM then PUT to GPU over and over (i.e. swapping)...
I was never directly editing the VRAM on the GPU, I was always having to PUSH to GPU

And I guess that's where the confusion was...

I kept saying "depending on the Application," "depending on the Application," me thinks some have too much gaming on the mind...

 

[cmaier]

Yes in MY APP, I load from SSD -> RAM, then when I "render" it gets transferred to the GPU.

i.e. My APP:

SSD -> RAM (edit/tinker/whatever) -> GPU [Intel chip]
now
SSD -> RAM (edit/tinker/whatever) [Apple chip] (no swap/update needed)

With the Apple it's all the same...

And yes in My App on Intel what was in MEMORY was MIRRORED on the GPU.
but then updated in RAM then PUT to GPU over and over (i.e. swapping)...
I was never directly editing the VRAM in GPU, I was always having to PUSH to GPU

And guess that's where the confusion was...

You shouldn‘t be doing that. That’s not how high performance graphics apps work.

 

[Jorbanead]

Here is another example. Why does a top of the line NVIDIA Titan have 12GB of VRAM?

Both the Titan and 3090 have 24 GB.

Yes in MY APP, I loaded from SSD -> RAM, then when I "render/rip/shred" it got transferred to the GPU. (ON INTEL, this was the case)

i.e. My APP:

SSD -> RAM (edit/tinker/whatever) -> GPU [Intel chip]
now
SSD -> RAM (edit/tinker/whatever) [Apple chip] (no swap/update needed)

With the Apple it's all the same...

And yes in My App on Intel what was in MEMORY was MIRRORED on the GPU.
but then updated in RAM then PUT to GPU over and over (i.e. swapping)...
I was never directly editing the VRAM on the GPU, I was always having to PUSH to GPU

And I guess that's where the confusion was...

I kept saying "depending on the Application," "depending on the Application," me thinks some have too much gaming on the mind...

What is your app? And also this explains a lot ?

 

[Mayo86]

In summary: Apple Silicon chips (so far at least) use unified memory and can attribute that memory to either CPU and/or GPU.

This was good entertainment.

 

[rezwits]

Anyone else here still confused by what OP means? Games and rendering apps are still applications. Why do you have 3 types of ram in your math?

Sorry OP it seems like you mean well.

You shouldn‘t be doing that. That’s not how high performance graphics apps work.

Right, High Performance Graphics Apps and Games for that matter, have say 1GB worth of Vector Data and Pixel (Texture) Data, in which that gets rendered out to the GPU in the form of JUST "PIXELS" say 4-6 GBs worth (in some cases).

Thanks Y'all...

 

[MacCheetah3]

That was a challenging train of thought to follow.

Anyway!..

To truly attempt to simplify and otherwise clarify with all due respect to @rezwits

But, if you are a user that requires 3D Graphics Memory, so for a rendering off a 3D graphic or Video Game world, then yes it's practically:
4GB + 12GB + 12GB = 28GB (of RAM & VRAM)

likewise an 8GB MacMini, would be 4GB + 4GB + 4GB = 12GB, if for Graphic Usage/Game Usage...

You cannot add that second 12. The limitation you are not realizing is...

The CPU needs to share that RAM

In other words @rezwits the remaining 12GB cannot be used to store both graphics and app data.

I will also try to simplify the following with a visual.

Supposedly I have a M1 Max system with 64GB of MEMORY (that can either be RAM or VRAM).
macOS occupies 4GB of MEMORY, so 60GB left for the user.
I open a BIG 3D project that needs 30GB of VRAM. Because M1 Max 64GB is Unified Memory, macOS assigns 30GB of MEMORY as VRAM to the program rendering that 3D project. Now I have 30GB of MEMORY to work with.

I need to open 30 tabs on Safari, 30 tabs on Microsoft Edge and 30 tabs on Google Chrome, which takes 12GB of MEMORY. Since those browsers don't render anything atm, they are considered as RAM. I have 18GB to work with now.

Suddenly I want to open a bunch of high res photos in photoshop for my work, also a few other video projects. Those combined used the remaining 12GB of MEMORY. macOS will give it to the program. Heck, swapping might already happening in the background.

Same task on a PC with a GPU that has 32GB of VRAM and 64GB of RAM. Now, instead of using 30GB of RAM, that 3D project rendering happens solely on GPU, using up all that 32GB of VRAM. However, the CPU cannot directly access those 32GB of VRAM, so some data will need to be transferred to the RAM for processing. This could eat a certain amount of RAM but probably not as much. The rest task happens mostly on RAM instead of VRAM.

EDIT: Did a quick change of the footnote.

My apologies on the crudeness.

Mind you that in the PC world they are introducing a new technology so CPU can access VRAM directly, bypassing RAM. If your logic holds any water, does that mean somehow my 64GB RAM + 32GB VRAM system now has 96GB of... RAM? VRAM? That doesn't make any sense.

Smart Access Memory and resizable BAR

Of course, the benefit is subjective:

AMD Smart Access Memory Tested, Benchmarked

Today we're taking a detailed look at how AMD's Smart Access Memory (SAM) technology influences performance in a wide range of games. All in all, we plan...

www.techspot.com

Nvidia Resizable BAR Tested, Benchmarked

Today we're testing the newly supported Resizable BAR feature on Nvidia GeForce GPUs by benchmarking an Asus ROG Strix RTX 3080 OC graphics card. How does it...

www.techspot.com

I have not seen or able to find productivity/creator-focused benchmarks (e.g., 3D rendering).

IDK tech changes maybe there is a NEWER one with MORE, IDK, just an example...

FYI, the RTX A6000 is Nvidia's current top workstation GPU with 48GB of VRAM. The MSRP is $4,650 USD (but with the current market, retailers are pricing it upwards of $7,000 USD).

 

[rezwits]

FYI, the RTX A6000 is Nvidia's current top workstation GPU with 48GB of VRAM. The MSRP is $4,650 USD (but with the current market, retailers are pricing it upwards of $7,000 USD).

dang thank you, and the thing is if you have an M1 MAX with 64GB and 32 GPU cores, you can essentially do the same thing as that 48GB card!

EXCEPT EXCEPT, we don't know the SPEED difference of the NIVIDA versus Apple's (meh) GPU implementation, I would gather a hunch tho that NVIDIA is "obviously" faster doing/using scientific applications.

BUT this was why I referenced "future" Quad Mac Pros which they are saying could have 256GB of UNIFIED RAM or IF You WANT GPU RAM!!!

I mean how much would an NVIDIA workstation GPU cost that had 128/256GB of VRAM?!?

 

[Eric40962005]

dang thank you, and the thing is if you have an M1 MAX with 64GB and 32 GPU cores, you can essentially do the same thing as that 48GB card!

EXCEPT EXCEPT, we don't know the SPEED difference of the NIVIDA versus Apple's (meh) GPUs implementation, I would gather a hunch tho that NVIDIA is "obviously" faster doing/using scientific applications.

BUT this was why I referenced "future" Quad Mac Pros which they are saying could have 256GB of UNIFIED RAM or IF CAN/WANT GPU RAM!!!

I mean how much would an NVIDIA workstation GPU cost that had 256GB of VRAM?!?

The RX A6000 is in a different galaxy than this M1 thing sorry man.

 

[cmaier]

The new title of this thread is even more specifically wrong than the old one.

 

[macintoshmac]

When you first look at the M1 Mac with Apple Silicon, you might say to yourself, if the 16GB of RAM is shared/pooled with the GPU and there is no swapping, then it’s like you have 32GB of RAM or RAM + VRAM. Which is not really true like some have said, but…

In some cases let’s take the 32GB RAM M1 Pro. You say well 32GB is like 64GB but it’s not right , it’s essentially (low balling) only 4GB for system and “some app” to “take advantage” of the remaning 28GB for GPU you theoretically get 28 + 28 + 4 = 60GB.

And that’s where I left off for a while, until I did the REAL MATH, you wanna hear the REAL MATH?

Go the other way:

Photo by Nana Dua on Unsplash

Let’s say you had a PC with 32GB of RAM GPU (workstation class), right? K, in order to LOAD UP that 32GB of VRAM on the GPU, you still have to have, in the Windows world 32GB + 4GBs or well really 48GB? Or 64GBs? + 32GB (hmm), but let’s say for sake of argument, you just went with, 32GB CPU + 32GB GPU, and only used 28GB of the CPU and 4GBs for your App that takes advantage of the 32GBs, and left 4GBs free on the GPU. How much RAM do you have? 64GB of 32 CPU (RAM) + 32 GPU (VRAM).

At this point you could actually say that 32GBs of Apple Silicon RAM is equivalent to 32GB CPU + 32GB GPU for a total 64GBs of RAM + VRAM.

BUT and here is the BIG BUT…

Let’s get to the swapping or lack there of.

On a PC if you’re swapping 24GBs you have to swap/load that into the GPU and that takes some time. But if it’s a Data Structure or complicated Data Set and you alter some of it via the CPU RAM, then on the next RENDER or whatever, sometimes you have to SWAP that whole 24GB chunk back in AGAIN to the GPU, (unless some kind of DMA, direct memory access) of the GPU RAM)!!

On the Mac with Apple Silicon if you edit the Data Structure or complicated Data Set it’s “just edited” and right there, don’t need FLUSH the GPU and reload all those voxels and textures etc, back into the GPU…

INSANITY…

So check this, out, as an aside…

If we get a QUAD Mac Pro Apple Silicon with 4 x M1 Max chips, with 64 GB of RAM, that would essentially be 250GB of VRAM!! Do you know what you can do with that? Do you know what you can do with that? And the fact that your app automatically inherits DMA? Because it’s shared/pooled memory? OMG

That’s like 512GBs of RAM+VRAM (or a GPU with 250GB of VRAM)

Silly NVIDIA card 32GB of RAM

Currently, there are NVIDIA Workstation Class GPUs with 32GBs of RAM, that cost $10,000!
That quad Mac Pro, could have 250GBs of VRAM and a true 128 GPU cores!!!

INSANE? Or like IDK what’s after INSANE?

Meanwhile the professionals that are professionals are just buying the machines and working on them, doing what they do best.