How much unified memory is accessible to the Apple GPU?

[ArPe]

Unified Memory Architecture is not new of course, it was used in SGI workstations (MIPS and Intel based) in the late 90s. To some extent AGP had the same aims, as do integrated graphics from Intel and AMD but with limits.

What are the memory limits on Apple GPU? I except that Apple Silicon Macs will improve the Application Saves States feature (from iOS) to make memory management more efficient and free up more memory for the foreground app.

If possible when you start reviewing your new Macs from next week please report how much memory usage you are seeing by the GPU - there are some creative apps and games which report video memory consumption. This will help users make better buying decisions if they need the full 16GB memory currently offered.

Thanks ?

 

[leman]

All of it, obviously. What would be the point otherwise? Also, the entire concept of “Video memory” becomes irrelevant.

 

[t0mat0]

I’d love to hear from Anandtech on what they think is possible from a larger M1X
(Assuming M1 roughly an A14X already). How much RAm can be close in the chip? Does it need to be on the same side?
If the chip is twice the height and width, Memory all around edges, could you get another 32-48GB to get a Mac up to 64GB?

Could you have more unified memory but off chip? Separate from the M1 or M1X?

the limitations of the Mac mini, 13” M1 MBP come from limitations of the M1 -
- 2 Thunderbolt ports
- 16GB RAM max

Hopefully soon we’ll see if we need to rethink how much RAM we need in an Apple Silicon Mac.

Going to be a classic “do I wait and get the better AS Mac, or buy now?” situation -
the new AS Mac lined up is a 24” iMac likely to be an M1 Mac - so we don’t have a ETA on higher than M1 Macs.

 

[ArPe]

All of it, obviously. What would be the point otherwise? Also, the entire concept of “Video memory” becomes irrelevant.

All of it? ?

The OS, apps and content are always dynamically adjusting how much memory they use. Swaps and scratches move in and out of memory to and from disk. Saved states move in and out of memory. Then you have some left over for the graphics.

Once people can show how much memory is being use by their apps and how it improves or impacts performance people will have a better idea.

I’ve always wanted system memory, video memory and storage to be one big persistent ultra fast memory pool. Everything instantly available, no boot times, no load times, no launch times. I’m hoping this implementation of UMA is a big step closer to that goal.

 

[ArPe]

I’d love to hear from Anandtech on what they think is possible from a larger M1X
(Assuming M1 roughly an A14X already). How much RAm can be close in the chip? Does it need to be on the same side?
If the chip is twice the height and width, Memory all around edges, could you get another 32-48GB to get a Mac up to 64GB?

Could you have more unified memory but off chip? Separate from the M1 or M1X?

the limitations of the Mac mini, 13” M1 MBP come from limitations of the M1 -
- 2 Thunderbolt ports
- 16GB RAM max

Hopefully soon we’ll see if we need to rethink how much RAM we need in an Apple Silicon Mac.

Going to be a classic “do I wait and get the better AS Mac, or buy now?” situation -
the new AS Mac lined up is a 24” iMac likely to be an M1 Mac - so we don’t have a ETA on higher than M1 Macs.

The die and SOC size is bound to be bigger on the forthcoming Pro machines, especially if they can increase GPU cores to something close to desktop GPU level. They will need 80 of those GPU cores to match the RTX 3080 according to the numbers being thrown around right now.

 

[leman]

The OS, apps and content are always dynamically adjusting how much memory they use. Swaps and scratches move in and out of memory to and from disk. Saved states move in and out of memory. Then you have some left over for the graphics.

Once people can show how much memory is being use by their apps and how it improves or impacts performance people will have a better idea.

I guess what I am trying to say is that this stuff is difficult to quantify. It's just... memory. There is no distinction between the CPU memory and the GPU memory. You just allocate as much as you need and you use it. And of course, you might run out of it if you more than the system can provide, but that is the beauty of the system — all memory is managed by the OS virtual memory system. If say, you are working with a large multi-layer photoshop file that needs tons of memory fo raster data, the OS might compress/offload application data that is not being currently used. If you then open back your browser, the OS will compress/offload some of the memory pages used by Photoshop. It's not like the GPU only gets the scraps — it is a full rights citizen in a system that comprises multiple different processor. If you want, it's literally communism — each processor contributes what it can and each device receives exactly what it needs to do it's job at the moment.

One can of course look at memory consumption figures and try to find some patterns, but in the end, what matters is "does it run well". If it does, what is the problem? Virtual memory management in a modern OS is a highly complex topic and it's very easy to find spurious patterns. If you are a regular user, common sense is more important.

Just my IMOs of course, I don't mean to demean your interest or anything like that.

 

[ArPe]

I guess what I am trying to say is that this stuff is difficult to quantify.

Not really. When the machines are in peoples hands and apps loaded with content or games are loaded up then a simple glance at Activity Monitor and in-app memory consumption will quantify how much memory is being used, how UMA improves efficiency (or not) and what other enhancements (App Saved States) in Big Sur are helping a foreground app performance.

These feedback and screenshots will be very helpful. The back end tech stuff is for the techies to debate about but most end users just want to know how much resource is being used so they can make a buying/upgrade decision.

This thread is just for that stuff, maybe should have posted in next week ??

 

[ArPe]

Bump. Today’s the day.

Please report how much memory you see the GPU utilizing. There are some apps, games and utilities that show this info.

Most importantly please check if there is a limit to how much memory the GPU can grab.

?

 

[t0mat0]

If you’re running 80 cores of Apple Silicon GOU, what hit to latency could an immaculately or Mac Pro take to have a larger unified memory cache and all those cores?
Could they go around the M1 in ever larger rectangles? Or have a fabric than ram to clusters of cores and memory?

 

[mrhick01]

Yes.