Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Do unified memory allow GPU have more VRAM?
- Thread starter sunny5
- Start date
- Sort by reaction score
It’s not so simple, as ram would still be needed by the system, but essentially yes.
Maxon (Cinema4D/Redshift) folks say the M1 Max can use a little over 40GB of RAM for the GPU, reserving the remaining RAM for CPU / System usage...?
Info towards this can be found in this post...
Info towards this can be found in this post...
Last edited:
Like others have said it's not as simple as yes as the RAM is a shared pool. But yes, essentially the GPU can in theory use it all. The biggest pro is not having to hold duplicate assets in both RAM and VRAM like a Windows PC so in theory a well optimized application will need less overall RAM.
400GB/s is actually a lot of bandwidth, this is approximately RTX 3070 levels of memory bandwidth, but unlike with a desktop gaming rig the CPU shares that bandwidth and gets to use that insane speed. Most desktops get ~30GB/s memory bandwidth from dual channel system RAM, so 400GB/s is quite a massive increase.
400GB/s is actually a lot of bandwidth, this is approximately RTX 3070 levels of memory bandwidth, but unlike with a desktop gaming rig the CPU shares that bandwidth and gets to use that insane speed. Most desktops get ~30GB/s memory bandwidth from dual channel system RAM, so 400GB/s is quite a massive increase.
Yes.
Not quite. You can always made a GPU with tons of private VRAM not shared with CPU.
The whole idea of needing VRAM is because the data path to the GPU is slow, so you use VRAM to hold data on the GPU side of the bottleneck. Once the bottleneck is removed, the need for VRAM goes away. Yes all of the RAM can be used for any purpose and the data never needs to move across a PCIe bus.
With this one I don’t think so. The GPU has access to all 64Gb regardless of what the CPU is doing. There is no ram partitioning like on Intel chips. UMA allows 100% access to both GPU and CPU at the same time.
Not quite: it’s true that there isn’t a strict partition but there does appear to be a limit on how much RAM the GPU can subscribe in order to ensure that the cpu has x amount of space for things it might need. So there are no duplications and the CPU and GPU can both see the entire RAM stack but the GPU has limits on how much it can use for solely itself. How this works in under the hood is probably pretty interesting.
Where have you seen this? Anything I can read/watch to learn more?
Last edited:
If I'm not wrong, macOS controls how much RAM is allocated for each process' 'VRAM' allocation. While Total War 3 Kingdoms may see 43GB, macOS itself would also need some 'VRAM' for its GUI rasterisation.
In any case, the concept of 'VRAM' is no longer relevant in macOS running in UMA Macs. Processes can ask for as much as macOS allows.
Any apps that can accurately monitor VRAM usage? just curious to see what happened when I enable the unlimited Vmem on TW 3 Kingdom.
All GPU accessible ratios point anywhere between 60-68.5% of the total ram available to the GPU.
I hope they allow a higher percentage for the Mac pros with more than one AS Soc..
I hope they allow a higher percentage for the Mac pros with more than one AS Soc..
Hmmm nothing comprehensive unfortunately, it’s just been noted that applications that say how much GPU Ram they can make use of it is always around 2/3 of the total. In contrast, I’ve not seen any such limit to a process normally. So in practice there is a limit to how much RAM a program can use dedicated to the GPU. However unclear exactly how this limit works or what happens if multiple programs are in use. I’ll see if I can dig up something more concrete.
Edit: @Boil ‘s post above linked to a post from @singhs.apps that’s pretty good, but I don’t think anyone has done a full technical analysis - at least not that I’ve found.
Last edited:
Most likely because CPU processes can make use of virtual memory, while GPU processes should not? I would think GPU processes should only be wired to physical memory space while the CPU can make use of the MMU.
Not that I know of tho. Not sure if macOS has APIs to allow queries of other processes memory usage.
I dunno ?♂️ maybe - it’s a good hypothesis. Unfortunately I don’t have one of my own test out and try to figure out the contours of these kinds of things. So there’s a limit to what I can say beyond reading what others have written. You know … it could also be an interesting question to ask of the Asahi Linux people when they finally get their graphics driver stack finished what limits if any there are on their end.
AFAIK there is no direct way querying how much memory does a metal device actually 'has', but the
MTLDevice protocol does offer a value called recommendedMaxWorkingSetSize which means 'An approximation of how much memory, in bytes, this device can use with good performance.' I doubt most programs use this to probe how much memory a Metal device can use and that does not equal to the total memory available to the system but a practical upper limit that a Metal device can use without performance penalty.
Last edited:
Interesting … however this apparent limit seems to exist in opencl too though as has been noted opencl is just a metal wrapper these days.
Edit: I’m not familiar with Metal but I know that openCL and CUDA have the ability to create different memory pools as device and host even in a “unified memory” programming model … can someone try the same in macOS and fill up the RAM with “GPU memory”?
Last edited:
A quick and dirty test showed that I can have an allocated size even larger than the physical ram by creating zero-filled buffers. The max buffer length on M1 is 8GB so I created 3 with 24GB total, and at least the playground does not crash on my 16GB M1 Mac, and the
MTLDevice indeed reported 24GB has been allocated. Maybe we could test again with MTLHeap and actually fill it with something not zero to see if the result differs.Regarding “maximal available memory”, there is no limit. Metal has a parameter returning “maximal memory you can use with good performance” which is probably the 43GB people have reported, but it’s just a guideline. You can allocate and use much more than that, it’s just that you might run into some swapping along the way.
@Gnattu: allocating is not enough, buffers are allocated lazily, you actually have to use them to have them pinned in memory. It’s the same on Intel Macs. You can allocate whatever you want with Intel or AMD GPUs but if you say start a compute kernel that bind buffers of total size larger than the GPU can handle, the app will crash. I expect no crashes on Apple Silicon, just swapping. Someone should test it. Still waiting for my Max ?
@Gnattu: allocating is not enough, buffers are allocated lazily, you actually have to use them to have them pinned in memory. It’s the same on Intel Macs. You can allocate whatever you want with Intel or AMD GPUs but if you say start a compute kernel that bind buffers of total size larger than the GPU can handle, the app will crash. I expect no crashes on Apple Silicon, just swapping. Someone should test it. Still waiting for my Max ?