Gaming on the M1

[the8thark]

Firstly I'd like to share a a few videos from youtube where a bloke tested a number of games on M1 (Macbook Pro). So the Mini should get similar performance and the iMac, slightly better performance (I assume).

The first video has 25 games tested which are a range of 64-bit Metal 2 and OpenGL games, iPhone & iPad apps on Mac, and 32-bit games.

Spoiler: click here to see the list of most of the 25 games in the video

Borderlands 3
Half Life 2
World of Warcraft
Shadow of the Tomb Raider
Minecraft
DOTA 2
Bioshock 2: Remastered
Among Us
F1 Mobile Racing
Stardew Valley (iOS)
Dead Trigger 2
Soul Knight
Monument Valley
Stardew Valley (Steam)
DiRT Rally
Borderlands 2
Counter-Strike: Global Offensive
Cuphead
Middle-earth: Shadow of Mordor
Dying Light
Hades
The Pathless
Deus Ex: Mankind Divided
Batman: Arkham City
The Sims 4

The second video has 33 games tested which are a range of arm64, Parallels 16, Rosetta 2, and CrossOver games.

Spoiler: click here to see the list of most of the 33 games in the video

Alien Isolation
A Plague Tale: Innocence
BioShock Remastered
The Bard’s Tale IV: DC
CS GO
SteamVR
GTA IV
The Elder Scrolls Online
SUPERHOT: MCD
DARK SOULS: REMASTERED
Thief
Crusader Kings III
CARRION
LEGO: The Incredibles
Fallout 4
Factorio
F1 2016
Pathfinder: Kingmaker - EPE
Moto Racer 4
Celes
Aragami
Outlast
The Sims 2
PAYDAY 2
Yooka-Laylee
The Long Dark
Skul: The Hero Slayer

The third video video has 32 games tested

Spoiler: click here to see the list of most of the 32 games in the video

Metro Exodus
Path of Exile
Spacebase Startopia
NBA 2K21 Arcade Edition
Yakuza 0
GWENT: The Witcher Card Game
Transport Fever 2
GZDoom
Wreckfest
Ori and the Blind Forest: Definitive Edition
Ori and the Will of the Wisps
Unreal Tournament
Frostpunk
Beyond Blue
The Last Campfire
DEVOTION
The Talos Principle
Hellblade: Senua's Sacrifice
Hitchhiker
Enter the Gungeon
Valheim
Intruder
Assassin's Creed IV: Black Flag
Disco Elysium: The Final Cut
Gravia
Cities: Skylines
Layers of Fear 2
Narita Boy
Little Nightmares

None of these games are bleeding edge or 2021 end graphics but the two videos here should give you a good overview of how capable the M1 is for gaming. It's much more capable than I first thought.

 

[dmccloud]

Using "conventional" logic, the M1's graphics performance shouldn't be nearly as good as it is. However, that is based on viewing things through the x86 (Intel) paradigm, where iGPUs are hot garbage at best, completely unusable otherwise. UMA plays a big role in this, since both the CPU and GPU can access the same data simultaneously instead of having to copy the data into both the system and iGPU partitions then reconciling the changes made on both copies. Think of it as editing a paper - if you have two people editing it separately, you then have to take the two edited copies and combine them into one document before publishing. However, if both people are editing side by side, their output is the finished product and can be published without the reconciliation stage, which improves overall turn time and lowers time to publication.

 

[Lowhangers]

The only game I play on my Mac mini is World of Warcraft. And it runs quite well add a respectable resolution at 60 frames per second on my M1 Mac mini. I have no complaints whatsoever, for the price I paid for the mini.

Apple has done a fantastic job with the M1, and I can only imagine how much better it’s going to get over the years as they improve it. I can’t imagine where webwill be in four or five generations of these processors.

 

[haralds]

Game performance is surprisingly good even for Windows games running under Windows 10 hypervisor and x86 emulaiton.

 

[George Dawes]

Imagine an appletv with the m1 chip !

Then they truly could call the Apple Arcade , arcade / console level ..

 

[leman]

Using "conventional" logic, the M1's graphics performance shouldn't be nearly as good as it is.

If by conventional logic you mean the popular belief "iGPUs are slow". It's the same as "ARM is slow" — just because a paradigm has been traditionally used for low-cost, low-performance devices, it doesn't mean that it cannot be used to build a high-performance device.

M1 has excellent potential for gaming performance, especially if games would use its unique features. Apple's TBDR allows one to radically simplify and streamline the code and resource usage for many modern rendering approaches. Traditional GPUs require you to jump through non-obvious hoops when doing some things, with Apple's devices is can be a walk in the park instead.

However, that is based on viewing things through the x86 (Intel) paradigm, where iGPUs are hot garbage at best, completely unusable otherwise. UMA plays a big role in this, since both the CPU and GPU can access the same data simultaneously instead of having to copy the data into both the system and iGPU partitions then reconciling the changes made on both copies.

UMA is much less important for games (than say, for content creation software), because it is very rare for games to require both CPU and GPU access to the data. The overwhelming majority of games use the CPU as a staging area for setting up resources that are then copied and used on the GPU. That's it.

And few more notes:

- Just the fact that you have UMA does not mean that you don't have to copy the data. Zero-copy functionality requires explicit use of specialized APIs and needs you to set up your data in very particular way. Games just don't bother with it. So you still have to copy the data, it's just the copy itself is much faster (you get full 68+GBps bandwidth instead of 15GBps of PCI-e 3.0 16x used in high-end gaming GPUs)

- Intel (and AMD) iGPUs have been using UMA for years. Their memory topology is practically identical to Apple's (CPU and GPU access memory via a shared last level cache). Apple simply has much more capable memory controllers (and also desktop-level GPU cache sizes) — they optimize for performance and efficiency here and not cost like other vendors.

 

[Spindel]

- Intel (and AMD) iGPUs have been using UMA for years. Their memory topology is practically identical to Apple's (CPU and GPU access memory via a shared last level cache). Apple simply has much more capable memory controllers (and also desktop-level GPU cache sizes) — they optimize for performance and efficiency here and not cost like other vendors.

No intel and amd iGPUs do not use UMA at all.

Intel and AMD iGPU sections of a part of the system RAM to be used as vRAM the rest of the system can not access this part of the RAM at all only the GPU can.

This is not the case with UMA, with UMA the entire system has access to all the RAM at all time.

And regarding games "not making use of UMA" it has nothing to do with the games and everything to do with the graphics drivers. The game just calls via API to the graphics driver, the driver then sorts memory handling.

 

[leman]

No intel and amd iGPUs do not use UMA at all.

Intel and AMD iGPU sections of a part of the system RAM to be used as vRAM the rest of the system can not access this part of the RAM at all only the GPU can.

This is simply wrong. Intel documentation clearly states that their iGPUs starting with at least Sandy Bridge (if I remember correctly) use UMA and they expose zero-copy CPU/GPU data extensions to OpenCL that work identically to zero-copy functionality of Apple Metal.

And regarding games "not making use of UMA" it has nothing to do with the games and everything to do with the graphics drivers. The game just calls via API to the graphics driver, the driver then sorts memory handling.

This is again wrong (at least in regards to the topic). Graphical APIs require you to manually copy the data (either by providing a data pointer upfront or by using copy/mapping APIs). So UMA or not UMA, if you are using something like MTLBuffer.contents() or MTLTexture.replace() you are always doing a memory copy from the storage allocated by your application to the storage allocated by the driver. You need to use MTLDevice.makeBuffer(bytesNoCopy: <...>) if you want zero-copy behavior.

To sum it up: just because your system has UMA, it does not mean that all data copies can be magically eliminated.

 

[dmccloud]

This is simply wrong. Intel documentation clearly states that their iGPUs starting with at least Sandy Bridge (if I remember correctly) use UMA and they expose zero-copy CPU/GPU data extensions to OpenCL that work identically to zero-copy functionality of Apple Metal.



This is again wrong (at least in regards to the topic). Graphical APIs require you to manually copy the data (either by providing a data pointer upfront or by using copy/mapping APIs). So UMA or not UMA, if you are using something like MTLBuffer.contents() or MTLTexture.replace() you are always doing a memory copy from the storage allocated by your application to the storage allocated by the driver. You need to use MTLDevice.makeBuffer(bytesNoCopy: <...>) if you want zero-copy behavior.

To sum it up: just because your system has UMA, it does not mean that all data copies can be magically eliminated.

Regardless of what Intel calls it, it is still not true UMA. Go into Best Buy and look at any of the laptops on display with only integrated graphics. You will see that they will say something like "8GB RAM (6.03 GB available). That's because they are still partitioning the physical RAM into virtual CPU/iGPU partitions. Intel likes to mislead or even lie in their marketing and sometimes technical materials. When they announced their 11th gen CPUs, they only compared their iGPU (which was actually just a rebranding of the "UMA" graphics used in the 9th and 10th gen parts) to AMDs Vega 4 and the nVidia MX250 (which had already been replaced by the MX350) in a bid to prove that their graphics were superior. A group of us watched that presentation at work and ripped Intel to shreds for the amount of misleading statements they crammed into a 47-minute prerecorded presentation.

 

[Ethosik]

The M1 beats or competes with the GTX 1050 Ti which is still #2 in the Steam hardware survey.

Steam Hardware & Software Survey

I really hope by the time it gets to the big boy Mac Pro, that the GPU is quite competitive. Trying to get even an RTX 3060 is impossible.

 

[Ethosik]

This is simply wrong. Intel documentation clearly states that their iGPUs starting with at least Sandy Bridge (if I remember correctly) use UMA and they expose zero-copy CPU/GPU data extensions to OpenCL that work identically to zero-copy functionality of Apple Metal.

If you read the document, it needs to be a programming change. How many software have this change in place?

Getting the Most from OpenCL™ 1.2: How to Increase Performance by...

Downloads Download Getting the Most from OpenCL™ 1.2: How to Increase Performance by Minimizing Buffer Copies on Intel® Proce

software.intel.com

To create zero copy buffers, do one of the following:

Use CL_MEM_ALLOC_HOST_PTR and let the runtime handle creating a zero copy allocation buffer for you

If you already have the data and want to load the data into an OpenCL buffer object, then use CL_MEM_USE_HOST_PTR with a buffer allocated at a 4096 byte boundary (aligned to a page and cache line boundary) and a total size that is a multiple of 64 bytes (cache line size).

When reading or writing data to these buffers from the host, use clEnqueueMapBuffer(), operate on the buffer, then call clEnqueueUnmapMemObject(). This paper contains code samples to demonstrate the best known practices on Intel® platforms.

 

[JouniS]

I expect that the M1 will remain good for the next year or two. By the time the component shortage ends, plenty of people will have current-generation consoles and GPUs, and new games will be designed to take advantage of the faster hardware. A 16-core M1 GPU would then be low end and 32 cores would be midrange.

There is always an intentional jump in hardware requirements some time after the release of a new console generation. If old consoles were still adequate for new games, there would be little reason to buy new models. On the other hand, because the time between console generations is ~7 years, any hardware that is good after the jump will remain good for many years.

 

[leman]

Regardless of what Intel calls it, it is still not true UMA. Go into Best Buy and look at any of the laptops on display with only integrated graphics. You will see that they will say something like "8GB RAM (6.03 GB available). That's because they are still partitioning the physical RAM into virtual CPU/iGPU partitions.

That's just how the driver model works. An obviously some memory is reserved for GPU purposes. It's not that different with M1: you can query the amount of available GPU memory and the number you get is not the total system RAM. That is just a reporting/"accounting" question and has little to do with UMA itself. Unified memory means that the GPU and CPU are able to physically (and logically) access the same memory in a cache-coherent way, and any reasonable model Intel iGPU has that capability.

Intel likes to mislead or even lie in their marketing...

Oh, that we absolutely can agree upon. Intel marketing likes to be, how to put it... "optimistic". But from the technical standpoint, Intel GPUs are quite interesting and not unsophisticated devices. They have very flexible compute capabilities.

If you read the document, it needs to be a programming change. How many software have this change in place?

Exactly the same thing for Apple Silicon. If you want zero copy behavior, you have to use a specific API and make sure that you allocate your memory in a particular way.

 

[LeeW]

The only game I play on my Mac mini is World of Warcraft. And it runs quite well add a respectable resolution at 60 frames per second on my M1 Mac mini. I have no complaints whatsoever, for the price I paid for the mini.

Apple has done a fantastic job with the M1, and I can only imagine how much better it’s going to get over the years as they improve it. I can’t imagine where webwill be in four or five generations of these processors.

This is me also, I only play WoW casually and dip in and out of Retail/Classic when time permits. Played since 2006 but don't raid any more just enjoy some questing and wandering around the world.

Runs amazing on the M1 Mac Mini, in comparison to my previous 2014 and 2018 devices. 45-60 FPS on a 3440x1440 @ max settings with a slight adjustment to shadows, it barely gets warm and handles it just fine.

 

[Ethosik]

Exactly the same thing for Apple Silicon. If you want zero copy behavior, you have to use a specific API and make sure that you allocate your memory in a particular way.

That is not how it works. Apple Silicon has access to any and all of the 16GB of RAM. There is no separate VRAM, but there still is for Intel iGPU.

 

[Hessel89]

Just based on geekbench scores I was expecting similar performance to an AMD RX 560.
This seems to be exactly the case in most games, with the only difference that I can set the drawing distance to it's absolute max on the M1. (thanks to the 16GBs of unified ram)

With Metal 2 enabled games it's a whole different ballgame though!
I can play X-plane 11 at 2560 x 1440 with everything set to Ultra / High and I'm still getting a solid 30 to 40 fps.
That's actually better then what a friend of mine is getting with his RTX 2060. All the while staying completely quiet!

It shouldn't be this powerfull but somehow it is..

 

[leman]

That is not how it works. Apple Silicon has access to any and all of the 16GB of RAM. There is no separate VRAM, but there still is for Intel iGPU.

I am not aware of any memory addressing limitations on Intel iGPUs. Both Intel and Apple devices access system memory via the same last level cache and the same memory controller. All Intel GPU memory is visible and cache-coherent with the CPU and via versa. Both Intel and Apple GPU devices are able to directly access system memory allocations (zero-copy) using opt-in APIs.

Any "VRAM size" reports etc. you might see are driver or OS-specific limitations. As far as I know, Intel GPU driver can adjust these lists dynamically, Apple Silicon uses very high natural limits from the start. Both do reserve certain amount of RAM for GPU purposes.

 

[Spindel]

I am not aware of any memory addressing limitations on Intel iGPUs. Both Intel and Apple devices access system memory via the same last level cache and the same memory controller. All Intel GPU memory is visible and cache-coherent with the CPU and via versa. Both Intel and Apple GPU devices are able to directly access system memory allocations (zero-copy) using opt-in APIs.

Any "VRAM size" reports etc. you might see are driver or OS-specific limitations. As far as I know, Intel GPU driver can adjust these lists dynamically, Apple Silicon uses very high natural limits from the start. Both do reserve certain amount of RAM for GPU purposes.

Yes it is true that the vram size can be adjusted dynamically with intel igpus.
But you don’t get the point, the alocated vram can still not be accessed by the rest of the system. So if you have a graphical entity that needs positioning (done by the CPU) that entity has to reside in both vram and system ram. Because the cpu needs to do its calculations on it and since it can’t access vram it needs a copy in system ram (all of this grossly simplified)

with UMA both the cpu and gpu can access the same memory adress.

 

[leman]

But you don’t get the point, the alocated vram can still not be accessed by the rest of the system. So if you have a graphical entity that needs positioning (done by the CPU) that entity has to reside in both vram and system ram. Because the cpu needs to do its calculations on it and since it can’t access vram it needs a copy in system ram (all of this grossly simplified)

This is simply incorrect. There is no "VRAM" with modern Intel or AMD iGPUs. The CPU and the GPU can access the same physical memory. As I wrote before, the CPU and the GPU share the caches and the memory controllers.

If you don't believe me, look at the available Vulkan memory pools for these devices (via vkGetPhysicalDeviceMemoryProperties). They only memory pools they offer is accessible by both the device (GPU) and the host (CPU).

with UMA both the cpu and gpu can access the same memory adress.

Precisely. Which is why I am puzzled that you would claim that Intel is not UMA. That's exactly how Intel iGPUs work.

 

[dmccloud]

This is simply incorrect. There is no "VRAM" with modern Intel or AMD iGPUs. The CPU and the GPU can access the same physical memory. As I wrote before, the CPU and the GPU share the caches and the memory controllers.

If you don't believe me, look at the available Vulkan memory pools for these devices (via vkGetPhysicalDeviceMemoryProperties). They only memory pools they offer is accessible by both the device (GPU) and the host (CPU).



Precisely. Which is why I am puzzled that you would claim that Intel is not UMA. That's exactly how Intel iGPUs work.

With Intel iGPUs, the RAM is split into two virtual partitions: one that is used as system RAM, the other is dedicated to the iGPU. So an HP or Dell laptop that ships with 8GB of RAM will only have between 6 and 7.5GB of RAM available for programs because the rest is allocated solely to the iGPU. Furthermore, data written to the system partition can only be accessed by the CPU, and data written to the GPU partition only accessible by the iGPU. This is why Intel's use of the term "UMA" is misleading at best. With the M1, Apple has granted full access to the entire memory address space to both the CPU and GPU cores. For applications where both the GPU and CPU are used (many MMOs, including World of Warcraft fit this criteria), data has to be copied into both partitions on the Windows side, then the changes reconciled before any output is carried out. With the M1, both the CPU and GPU cores can access and modify the data simultaneously. This speeds up overall performance because the reconciliation step is no longer needed.

 

[leman]

With Intel iGPUs, the RAM is split into two virtual partitions: one that is used as system RAM, the other is dedicated to the iGPU. So an HP or Dell laptop that ships with 8GB of RAM will only have between 6 and 7.5GB of RAM available for programs because the rest is allocated solely to the iGPU. Furthermore, data written to the system partition can only be accessed by the CPU, and data written to the GPU partition only accessible by the iGPU. This is why Intel's use of the term "UMA" is misleading at best.

Where do you have this information from? This is contrary to all technical documentation and API behavior that I am aware of. As to "virtual partition" — that's just the RAM reserved for the GPU use. Apple systems also do that.

 

[Ethosik]

Where do you have this information from? This is contrary to all technical documentation and API behavior that I am aware of. As to "virtual partition" — that's just the RAM reserved for the GPU use. Apple systems also do that.

Again, its a programming change. Running an old game from 2005 on a modern Intel iGPU will still need to perform a copy to the iGPU memory. You need to actually state in the code that you want zero memory buffer.

Since Apple doesn't have a dedicate iGPU VRAM, this issue does not occur. It does not have this separation.

Please show where Apple systems show dedicated sections of RAM ONLY for GPU. I have a Mac mini M1 here and I cannot find where it states only 1 or 2 GB of VRAM available.

 

[leman]

Again, its a programming change. Running an old game from 2005 on a modern Intel iGPU will still need to perform a copy to the iGPU memory. You need to actually state in the code that you want zero memory buffer.

And again, it’s the same thing when you use Metal or any other GPU API. I am really confused as to what you are trying to say.

Please don’t take it the wrong way, but I just have to ask at this point: what is your experience with GPU programming? It doesn’t sound to me like you are too familiar with graphical APIs.

Since Apple doesn't have a dedicate iGPU VRAM, this issue does not occur. It does not have this separation.

Of course it does. Please refer to the Metal documentation. All basic API assumes that you are copying data into driver-allocated memory buffers. If you want to achieve zero-copy functionality, you need to use specialized APIs (as you say, “it’s a programming change”). I think it’s like the fifth time I’m mentioning it. It’s getting a bit frustrating.

Please show where Apple systems show dedicated sections of RAM ONLY for GPU. I have a Mac mini M1 here and I cannot find where it states only 1 or 2 GB of VRAM available.

“VRAM available” is just some information that driver/the OS gives you. Intel iGPU don’t have any “VRAM”. They have system memory managed by the GPU driver. Just like Apple. The rest is just driver behavior. An Intel driver on Windows might limit your maximal GPU RAM usage to 4GB or so (honestly, I wouldn’t know), Apple is more flexible.

 

[Ethosik]

Please don’t take it the wrong way, but I just have to ask at this point: what is your experience with GPU programming? It doesn’t sound to me like you are too familiar with graphical APIs.

I create video games, work with pixel and vertex shaders in Direct X, XNA, MonoGame, Unity, Unreal and more. I have my experience which is why several people that have posted saying Intel doesn't have this issue do NOT know its a programming change - and many people did not even realize it was a programming change. I have not been able to find an example that you keep claiming that Apple's M1 still suffers from this issue, and I have searched for days. I am not able to find the flags needed for zero memory buffer like you do for Intel. So please, point me in the right direction instead of just saying it 5 times. If you have the proof, I would definitely like to read it. I have NOT seen ANY proof that the M1 still has its own dedicated GPU memory allocation. EVERYTHING I read and ALL documentation I have seen states the M1 GPU can access ANY and ALL of the 8GB or 16GB of the Unified Memory.

I have seen dozens of developer discussions and developer opinions like this one:

Why The Apple M1 Chip Is So Fast - A Developer Explains | Production Expert

You may have wondered why the Apple M1 Chip is so fast? In this article, we learn from a developer as to what makes, even these entry-level Macs, so fast.

www.pro-tools-expert.com

With the M1, this is also part of the SoC. The memory in the M1 is what is described as a ‘unified memory architecture’ (UMA) that allows the CPU, GPU, and other cores to exchange information between one another, and with unified memory, the CPU and GPU can access memory simultaneously rather than copying data between one area and another.

I have tried to look at the metal documentation, and I do NOT see where a programming change is required like it is for Intel iGPU. Again, if you have a specific part of the document, please point me to it. But all these accounts from developers, and nothing I have seen thus far points me to the same issue that Intel IGPU has.

And you do know that when we say VRAM in regards to Intel iGPU, we actually mean "segmented memory allocated to the GPU by the driver". VRAM is short and easy and still accomplishes the conversation.

 

[leman]

I create video games, work with pixel and vertex shaders in Direct X, XNA, MonoGame, Unity, Unreal and more.

Ah, that's great, so we can discuss this on a technical level.

Edit: sorry, the forum inserts emojis into the URLs automatically :/

I have my experience which is why several people that have posted saying Intel doesn't have this issue do NOT know its a programming change - and many people did not even realize it was a programming change. I have not been able to find an example that you keep claiming that Apple's M1 still suffers from this issue, and I have searched for days. I am not able to find the flags needed for zero memory buffer like you do for Intel.

If you want to have zero-copy behavior, that is, use your own memory allocation as GPU buffer you have to explicitly use makeBuffer(bytesNoCopy:lengthptions:deallocator API, which is subject to the same alignment restrictions as Intel zero-copy extensions. If you want to hav zero-copy textures, you have to use the above API and then allocate texture from this buffer.

The regular APIs (e.g. makeBuffer(lengthptions, makeTexture(descriptor ) etc. have the driver allocate opaque memory, which you then need to populate yourself. This is how the API is used normally, which means that you have to copy memory around even on an Apple Silicon UMA system.

I have NOT seen ANY proof that the M1 still has its own dedicated GPU memory allocation. EVERYTHING I read and ALL documentation I have seen states the M1 GPU can access ANY and ALL of the 8GB or 16GB of the Unified Memory.

M1 obviously does not have have any separate GPU memory allocation since t does not have any separate GPU memory. Neither do modern Intel or AMD iGPUs. It seems to me you might be mixing up VRAM and driver-managed memory. For example, if you call makeTexture(descriptor, the driver will allocate and manage the memory for you, and you won't even be able to get a pointer to that allocation. If you are using a Mac with a dedicated GPU, that memory could reside in GPU VRAM only (private resource type) or have DMA-synchronized copy in the system RAM (managed resource type). On a UMA device (e.g. Apple Silicon, Intel iGPU), this will be located in system memory.

And you do know that when we say VRAM in regards to Intel iGPU, we actually mean "segmented memory allocated to the GPU by the driver". VRAM is short and easy and still accomplishes the conversation.

But this "VRAM" is nothing more than a portion of the system memory managed by the driver. All of this memory resides in the regular physical RAM and is accessed through the same cache and memory controller as the CPU uses, which means that CPU can transparently and cache-coherently access any RAM the GPU driver allocates.

This can all be illustrated neatly if we examine Vulkan memory pools for Intel iGPUs. For example, here: https://vulkan.gpuinfo.org/displayreport.php?id=11242#memory As you can see the GPU device has a single memory heap of 16GB (which I assume is the total RAM of the machine) and all of this memory is DEVICE_LOCAL and can be allocated as HOST_VISIBLE and HOST_COHERENT. In other words, that's just normal system RAM, with no magic partitioning or access restrictions.