Bad GPU performance at 4K

[Homy]

You said that my "question suggested that the amount was important in the discussion." But that was not my intent. What I said was that I would want to make valid comparisons. You say that "when you have enough or too much RAM available the amount doesn’t matter after a certain point," which is true.

However UMA RAM operation is very different, and probably evolving. E.g. you reference VRAM in your discussion of Tomb Raider benchmarking, yet VRAM does not exist on Apple silicon. So IMO we do not know (or at least I certainly don't) exactly how RAM allocates under gaming on M3 chips. We do know that the same RAM is used for everything, concurrently.

So if I was to make an effort at comparing my point is that I wanted to know what the OP's RAM was because - IMO - RAM is an important parameter to consider in any analysis of M series performance, and I would want to make valid comparisons. Your experience with game demands on VRAM that does not exist in M-series Macs is meaningful (and hella more than mine), but from an experimental standpoint I would not automatically assume that on M3 games will absolutely see no impact from having in excess of 36 GB UMA RAM.

Note that if the OP had responded "I have 128 GB" the point would be moot and I would have no concerns that my 96 RAM might affect comparative results. Also note that I respect your opinion and I expect that it is very likely correct, and that most likely more than 36 GB RAM would not be of value if the game is the only app running.

You have the right to be curious and wanting to do an accurate comparison. Yes, we all know that UMA works differently and the CPU/GPU has access to the whole physical memory. Yes, the RAM size is also an important factor but again it all comes down to what you’re using the computer for, if your specs are enough or if you would benefit from better specs.

VRAM does exist even in UMA and on Apple Silicon, just not as distinct as in the case of dedicated GPUs. According to Apple ”the limit on amount of GPU resources an app can allocate has two values to be aware of. The total amount of GPU resources that can be allocated, and more critically, the amount of memory a single command encoder can reference at any one time.” The ”recommended maximum working set size” command helps you to control how much memory is available/used. Metal is able to allocate more memory beyond this though.

Metal Compute on MacBook Pro - Tech Talks - Videos - Apple Developer

Discover how you can take advantage of Metal compute on the latest MacBook Pro. Learn the fundamental principles of high-performance...

developer.apple.com

I’m not familiar with how it’s used exactly, if it’s written in both macOS and applications but Apple states that the GPU working set size for 32 GB is 21 GB. That’s what my screenshot of Tomb Raider shows. With 64 GB you have 48 GB VRAM. So it’s not recommended to exceed these values in order to have a reliable system.

Games are just like other software. The data, pixels, shaders, effects and other stuff are loaded into the memory and displayed. The more memory you have the more data you can have access to without having to write to/read from the disk. That improves the performance. That’s why the highest resolutions like 4K and ultra settings require large VRAM. That’s how it works with UMA on Macs too. The difference is that thanks to UMA you have access to more shared memory and don’t have to swap data between memory and disk as much.

Now when you have a game like Deus Ex that needs about 8-9 GB for max resolution and settings you wouldn’t benefit at all from more free RAM because there’s nothing more to be loaded into the memory. You would absolutely see no impact on the performance from 96 GB. It’s like being hungry and feeling full after a burger. Another table full of pizzas and tacos wouldn’t matter because you’re not hungry anymore. If you find a game that would use 70 GB of RAM you would benefit from 96 GB but the largest GPU VRAM on PC side is 24 GB so that’s the maximum you need for modern games.

In fact the peak VRAM use even with ray tracing at 4K is 13.9 GB in Hogwarts Legacy. Even on PC Deus Ex uses only 5.5 GB VRAM.

Other applications can certainly benefit from large memory. ArtIsRight did a comparison between M2 Ultra with 64 GB and 128 GB. The performance in apps like Lightroom and Photoshop was almost identical. Only when he used a 56 GB file in Photoshop he could see a huge advantage with 128 GB. That’s pretty logical because with 64 GB not much is left for the system to work with.

 

[Nozgog]

Are you using the MacBook Pro plugged in and on „High Performance“ setting in System Preferences?

Do you have Crossover 23/24 with D3DMetal turned on from Apple?

Do you have the Retina setting in Crossover turned off? (Huge performance penalty)

V-Sync is a general performance problem

 

[Yebubbleman]

I find it frustrating that even with a maxed out M3 Max Pro, running games, either native or in Crossover, at native 4K seems to be impossible. Even older titles like Deus Ex Mankind Divided (Mac version) completely fall apart when I ramp the resolution up to 4K (getting 30-40 FPS on highest settings - this computer should absolutely devour a 2017 game). Same with not very demanding games, retro games running through Crossover, like Turbo Overkill - works fine at 1440p but 4K is not really playable.

Unless you're running a game optimized for Metal (at the bare minimum), you're not going to get great performance out of any Apple Silicon Mac for gaming. Having said game be Apple Silicon native and not running translated via Rosetta 2 really does help too.

But that's really the rub; if modern games are not optimized and built around Apple Silicon Macs, then it will suck to use Apple Silicon Macs to play them.

 

[loui100]

Are you using the MacBook Pro plugged in and on „High Performance“ setting in System Preferences?

Do you have Crossover 23/24 with D3DMetal turned on from Apple?

Do you have the Retina setting in Crossover turned off? (Huge performance penalty)

V-Sync is a general performance problem

I run Deus Ex native version, but as for Crossover, yes v24, D3DMetal. I use Retina (High Resolution) setting - all it does is allow you to run at higher resolutions. There is a huge performance penalty on account of higher resolutions (otherwise it runs at 1080p), I don't think it causes any additional performance strain than normally setting a higher resolution?

 

[loui100]

You have the right to be curious and wanting to do an accurate comparison. Yes, we all know that UMA works differently and the CPU/GPU has access to the whole physical memory. Yes, the RAM size is also an important factor but again it all comes down to what you’re using the computer for, if your specs are enough or if you would benefit from better specs.

VRAM does exist even in UMA and on Apple Silicon, just not as distinct as in the case of dedicated GPUs. According to Apple ”the limit on amount of GPU resources an app can allocate has two values to be aware of. The total amount of GPU resources that can be allocated, and more critically, the amount of memory a single command encoder can reference at any one time.” The ”recommended maximum working set size” command helps you to control how much memory is available/used. Metal is able to allocate more memory beyond this though.

Metal Compute on MacBook Pro - Tech Talks - Videos - Apple Developer

Discover how you can take advantage of Metal compute on the latest MacBook Pro. Learn the fundamental principles of high-performance...

developer.apple.com

I’m not familiar with how it’s used exactly, if it’s written in both macOS and applications but Apple states that the GPU working set size for 32 GB is 21 GB. That’s what my screenshot of Tomb Raider shows. With 64 GB you have 48 GB VRAM. So it’s not recommended to exceed these values in order to have a reliable system.

View attachment 2360412

Games are just like other software. The data, pixels, shaders, effects and other stuff are loaded into the memory and displayed. The more memory you have the more data you can have access to without having to write to/read from the disk. That improves the performance. That’s why the highest resolutions like 4K and ultra settings require large VRAM. That’s how it works with UMA on Macs too. The difference is that thanks to UMA you have access to more shared memory and don’t have to swap data between memory and disk as much.

Now when you have a game like Deus Ex that needs about 8-9 GB for max resolution and settings you wouldn’t benefit at all from more free RAM because there’s nothing more to be loaded into the memory. You would absolutely see no impact on the performance from 96 GB. It’s like being hungry and feeling full after a burger. Another table full of pizzas and tacos wouldn’t matter because you’re not hungry anymore. If you find a game that would use 70 GB of RAM you would benefit from 96 GB but the largest GPU VRAM on PC side is 24 GB so that’s the maximum you need for modern games.

In fact the peak VRAM use even with ray tracing at 4K is 13.9 GB in Hogwarts Legacy. Even on PC Deus Ex uses only 5.5 GB VRAM.

View attachment 2360415
View attachment 2360416

Other applications can certainly benefit from large memory. ArtIsRight did a comparison between M2 Ultra with 64 GB and 128 GB. The performance in apps like Lightroom and Photoshop was almost identical. Only when he used a 56 GB file in Photoshop he could see a huge advantage with 128 GB. That’s pretty logical because with 64 GB not much is left for the system to work with.

View attachment 2360417

Yeah, I have 48 GB RAM and never get any close to straining the memory limit (in the Metal graphs memory usage is always green).

 

[jeanlain]

Nor have I, but some are saying that it’s the best Mac port.

These people tend to confuse general optimization with the quality of a macOS port, and forget that some games with good render quality run well on PC/consoles with mediocre GPUs.
None of those claming that it is a good port has compared the Mac to the PC, so I wonder how they came to their conclusion.

In turns out that Death Stranding is not a very good port. Performance comparison with Windows puts the M1 Max behind the RTX 2060, although it's 30% faster in 3DMark Wild Life and has much higher FLOPs rating (+60%). So the Mac is at least 40% behind where it should be.
I wouldn't be surprised if the performance delta in games like RoTR (intel) was lower than in Death Stranding.

 

[diamond.g]

These people tend to confuse general optimization with the quality of a macOS port, and forget that some games with good render quality run well on PC/consoles with mediocre GPUs.
None of those claming that it is a good port has compared the Mac to the PC, so I wonder how they came to their conclusion.

In turns out that Death Stranding is not a very good port. Performance comparison with Windows puts the M1 Max behind the RTX 2060, although it's 30% faster in 3DMark Wild Life and has much higher FLOPs rating (+60%). So the Mac is at least 40% behind where it should be.
I wouldn't be surprised if the performance delta in games like RoTR (intel) was lower than in Death Stranding.

It appears like modern mac games that are in 3d seem to be relying on upscaling for performance (I see similar issue on PC side to be honest).

 

[throAU]

I find it frustrating that even with a maxed out M3 Max Pro, running games, either native or in Crossover, at native 4K seems to be impossible. Even older titles like Deus Ex Mankind Divided (Mac version) completely fall apart when I ramp the resolution up to 4K (getting 30-40 FPS on highest settings - this computer should absolutely devour a 2017 game). Same with not very demanding games, retro games running through Crossover, like Turbo Overkill - works fine at 1440p but 4K is not really playable.

4k is still "hard" for modern desktop hardware at high details and as impressive as the M1/M2/M3 series mobile GPU is, it is still a mobile graphics chip running inside of 100 watts for the entire SOC (CPU/GPU/RAM/SSD/etc.) - not 200-300+ watts for the GPU alone.

Apple silicon is truly impressive, but it isn't magic and high power GPUs need.... lots of power. Even if the M series is 2x as efficient as Nvidia or AMD (and its not that much more efficient), its still running in a far smaller power envelope.

You're simply not going to match the performance of a full-fat modern desktop GPU - and even those struggle running current things at max details at 4k without resorting to scaling hacks.

4k is a huge increase in performance over 1080p and especially on a laptop display, probably not worth it. Frame rate is more important.


The translation layers only make things worse - but even a fully optimised metal game is going to struggle vs. modern desktop hardware.

 

[diamond.g]

4k is still "hard" for modern desktop hardware at high details and as impressive as the M1/M2/M3 series mobile GPU is, it is still a mobile graphics chip running inside of 100 watts for the entire SOC (CPU/GPU/RAM/SSD/etc.) - not 200-300+ watts for the GPU alone.

Apple silicon is truly impressive, but it isn't magic and high power GPUs need.... lots of power. Even if the M series is 2x as efficient as Nvidia or AMD (and its not that much more efficient), its still running in a far smaller power envelope.

You're simply not going to match the performance of a full-fat modern desktop GPU - and even those struggle running current things at max details at 4k without resorting to scaling hacks.

4k is a huge increase in performance over 1080p and especially on a laptop display, probably not worth it. Frame rate is more important.


The translation layers only make things worse - but even a fully optimised metal game is going to struggle vs. modern desktop hardware.

1440P is a good middle ground.

It isn't clear to me if the promotion displays require vsync (like freesync) or if it has to be off (like gsync) to work correctly, otherwise having high framerates (90+) with screen tearing isn't great. Really what you want is 1% lows higher than 60fps.

 

[Ctrlos]

I find it frustrating that even with a maxed out M3 Max Pro, running games, either native or in Crossover, at native 4K seems to be impossible. Even older titles like Deus Ex Mankind Divided (Mac version) completely fall apart when I ramp the resolution up to 4K (getting 30-40 FPS on highest settings - this computer should absolutely devour a 2017 game). Same with not very demanding games, retro games running through Crossover, like Turbo Overkill - works fine at 1440p but 4K is not really playable.

Are you running on a laptop or larger display? Even on a 16" Macbook Pro you won't notice any difference between 1080p and 4k visual fidelity. Frame rate on the otherhand....

 

[diamond.g]

Are you running on a laptop or larger display? Even on a 16" Macbook Pro you won't notice any difference between 1080p and 4k visual fidelity. Frame rate on the otherhand....

You should notice a difference in texture quality and particle effect quality. Shouldn't need as much AA at higher resolutions. Edges of Shadows should look better. Certain transparency effects should have less noticeable dithering.

 

[Ctrlos]

You should notice a difference in texture quality and particle effect quality. Shouldn't need as much AA at higher resolutions. Edges of Shadows should look better. Certain transparency effects should have less noticeable dithering.

I mean I might stop and gawk at the scenery in Horizon Forbidden West or Kena: Bridge of Spirits but none of these are on the Mac so normally I'm moving too quickly to notice.

 

[loui100]

Are you running on a laptop or larger display? Even on a 16" Macbook Pro you won't notice any difference between 1080p and 4k visual fidelity. Frame rate on the otherhand....

27-inch 4K Display. Definitely can tell the difference between 1080p and 4K lol. It's just way, way sharper

 

[HiVolt]

Tiny AS GPU will never be a 4090. You need to start to accept it.

 

[throAU]

27-inch 4K Display. Definitely can tell the difference between 1080p and 4K lol. It's just way, way sharper

As above 1440p is a decent compromise. Either way, you'll probably want to make use of upscaling if possible to get better frame rate. If you're on an Mx-Pro, 1440p using upscaling is probably a decent target, if you're on a Max, maybe 4k using upscaling is feasible for reasonably modern titles.

I still use upscaling on my 6900XT system in my sig quite frequently to get better frame rate. And that's a massively more capable GPU than almost anything, if not everything apple makes, especially in raster performance (less so at RT).

like I said, apple silicon is great, but there are still limits to what is achievable in the thermal and power envelope apple is playing in, and its nowhere near modern desktop GPU hardware running in hundreds of watts.

There are edge cases, particularly in pro workloads where the unified memory helps a lot to level the playing field, especially in very large data sets where the unified memory can effectively give the GPU far more VRAM than PC alternatives - but gaming is not one of those scenarios - all the games are optimized to fit in typical PC VRAM limits.

 

[diamond.g]

As above 1440p is a decent compromise. Either way, you'll probably want to make use of upscaling if possible to get better frame rate. If you're on an Mx-Pro, 1440p using upscaling is probably a decent target, if you're on a Max, maybe 4k using upscaling is feasible for reasonably modern titles.

I still use upscaling on my 6900XT system in my sig quite frequently to get better frame rate. And that's a massively more capable GPU than almost anything, if not everything apple makes, especially in raster performance (less so at RT).

like I said, apple silicon is great, but there are still limits to what is achievable in the thermal and power envelope apple is playing in, and its nowhere near modern desktop GPU hardware running in hundreds of watts.

There are edge cases, particularly in pro workloads where the unified memory helps a lot to level the playing field, especially in very large data sets where the unified memory can effectively give the GPU far more VRAM than PC alternatives - but gaming is not one of those scenarios - all the games are optimized to fit in typical PC VRAM limits.

Do you think Apples RT performance (Metal RT vs DXR) is better than a 2080ti?

 

[throAU]

Do you think Apples RT performance (Metal RT vs DXR) is better than a 2080ti?

On which chip? M3-Max?

Haven't compared it but I'm sure there are benchmarks out there.

Either way they're two entirely different products - if you want to run a recent Mac, a 2080 or any other Nvidia card isn't an option.

 

[diamond.g]

On which chip? M3-Max?

Haven't compared it but I'm sure there are benchmarks out there.

Either way they're two entirely different products - if you want to run a recent Mac, a 2080 or any other Nvidia card isn't an option.

I meant as a performance comparison. IIRC the 6900XT in RT is as fast as a 2080ti (again in games, not HIP vs Optix). We don't really have metrics on M3 Max RT performance in games.

 

[throAU]

I meant as a performance comparison. IIRC the 6900XT in RT is as fast as a 2080ti (again in games, not HIP vs Optix). We don't really have metrics on M3 Max RT performance in games.

Depending on the game - in pure raster the 6900XT out-performs the 4090 in some titles (no RT)

 

[Irishman]

These people tend to confuse general optimization with the quality of a macOS port, and forget that some games with good render quality run well on PC/consoles with mediocre GPUs.
None of those claming that it is a good port has compared the Mac to the PC, so I wonder how they came to their conclusion.

In turns out that Death Stranding is not a very good port. Performance comparison with Windows puts the M1 Max behind the RTX 2060, although it's 30% faster in 3DMark Wild Life and has much higher FLOPs rating (+60%). So the Mac is at least 40% behind where it should be.
I wouldn't be surprised if the performance delta in games like RoTR (intel) was lower than in Death Stranding.

Both can be true at the same time, DS being the best Mac port, but being outperformed by certain Windows configurations. I could be picky, too, about RoTR and 3DMark being introduced without any acknowledgment that we’re in a transition period arguably more disruptive than going from Motorola to Intel.

And, finally, who’s to say what the Mac “should be” in relation to other games and benchmarks. Does the kink you provided show their methodology?

 

[jeanlain]

Both can be true at the same time, DS being the best Mac port, but being outperformed by certain Windows configurations.

Not by a configuration equipped with a GPU that is 25% slower in cross-platform benchmarks and has much lower specs.

I could be picky, too, about RoTR and 3DMark being introduced without any acknowledgment that we’re in a transition period arguably more disruptive than going from Motorola to Intel.

Except that 3DMark runs well on Macs compared to windows PCs.

And, finally, who’s to say what the Mac “should be” in relation to other games and benchmarks. Does the kink you provided show their methodology?

I'm sure they tested with the same settings, it would be misleading if they didn't. These guys know what they're doing.

OTOH, those claiming it is the best Mac port should come up better arguments than "this games looks good and runs well". Yeah, it runs well on a PS4 too.

 

[Irishman]

Not by a configuration equipped with a GPU that is 25% slower in cross-platform benchmarks and has much lower specs.

Except that 3DMark runs well on Macs compared to windows PCs.

I'm sure they tested with the same settings, it would be misleading if they didn't. These guys know what they're doing.

OTOH, those claiming it is the best Mac port should come up better arguments than "this games looks good and runs well". Yeah, it runs well on a PS4 too.

Digital Foundry do know their stuff, but I’ve watched through their whole video and haven’t seen any of the points that you mentioned. I saw more comparisons between the iPhone and Mac performance than anything on the PC side.

Are you able to recall a time stamp or a time span so that I don’t have to scrub through the whole thing again?

ETA: Any luck? Yes? No? Maybe?