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.
M5 gaming
- Thread starter M4pro
- Start date
- Sort by reaction score
Yeah sometimes they update articles, but unfortunately they settled on CB R23 and so don't always test R24, so maybe the same will be true for CP2077 - they will eventually test the native version and occasionally do so, but normally it will be non-native.
Not sure. In fact, given that I agree that it's a pretty demanding test in terms of bandwidth I would've expected more from this update since Apple pushed bandwidth higher. Looking at the Steel Nomad Light numbers, I suppose the A19 uplift in SNL was only a little better than the M5 - 38% over the A18 vs like 33% for the M5 over M4 in SNL, but still it's only like 28% for the M5 over the M4 for the full SN. Interesting.
EDIT: you know ... maybe SNL and especially SN don't do much FP16? That's a big part of the raster performance increase for the M5. If the engine doesn't do much of that, it would look less impressive. Then again, SNL iPhone performance improvement is close to the 40% we see for some of the better titles even for the M5, so 🤷♂️
Last edited:
My original point was that there is a juicy market out there for both Apple (more people buying Macs) and game publishers and developers (more people buying their game on Mac), if the situation improved with game compatibility.
ie, it's worth the effort from Apple and game publishers and developers to improve things.
You're saying Apple silicon game performance is a barrier at the hardware level, but you're citing performance over a translation layer like Crossover. Running games through translation layers isn't a valid performance metric to talk about how powerful the hardware is, because the mere fact you're running it non natively is likely to completely tank performance.
Your experience is still valid, ie if you want to run Bethesda games right now on Mac, that is the performance you will get, but it's not valid in the context of talking about how powerful Apple's chips are for gaming.
I should clarify as well that Crossover, Game Porting Toolkit, and other tools like that on Mac shouldn't be compared to Proton. All Proton has to do is translate from Windows to Linux, because the hardware is identical.
Crossover has to:
- translate from x86 to ARM
- translate from Windows to macOS
- translate from whatever graphics API is being used to Metal
Thanks for your input, I appreciate offering your opinion.
The market has largely been the same for 5+ years and there's been scant new titles.
If its worth it, why haven't we seen more games coming to macOS?
What I'm saying is not in some theoretical future but rather in 2025, if someone wants to play a AAA game, its more likely then not that its a windows only game, and thus he/she will need crossover, and with the added overhead, the M4, and M4 Pro are inadequate. Will that change in the future maybe, but as of today and to a large extent 2026 crossover is the a mac owners best bet to play a game
A little extra pedantry from me (sorry): For a lot of games, Proton will also often have to translate from DX12 to Vulkan - some are already Vulkan of course, but many (most?) are not. But another little translation quirk on the Mac is also translating from 4KB page sizes to 16KB. Not sure about performance impact of that one but it's another thing WINE-Mac/XOver has to do and Apple spent a lot of time ensuring that worked so user space and kernel space could have different page sizes specifically for that reason (also for older macOS titles and Rosetta 2).
Last edited:
These things take time. Especially when Macs have been physically bad for gaming for decades.
Fair enough but not really relevant to what I was talking about.
Edit: Although you really shouldn't say the M4 line is inadequate because it can't brute force through the translation layer penalty. It's not just some blanket percentage performance penalty that if it was better it could just push through. It's like putting the wrong fuel in a plane and criticising the plane engine for not being able to generate enough power to fly.
Maybe the plane is a bad purchase because it's too hard to get the right fuel, but you wouldn't be citing the rpm of the plane engine as the reason.
Last edited:
Looking back the the M5 announcement, they talk about the shader core being 30% faster. So maybe the Steel Nomad scores are more reasonable?
Edited thanks!
Maybe but then that would imply no additional benefit from bandwidth which ... shouldn't be the case? I dunno.
The Cinebench GPU R24 benchmark does show the kind of improvement we anticipated from M5 GPU hardware-accelerated raytracing.
(CB R24 tests GPU-rendering using Maxon Redshift).
Last edited:
I was actually talking about the CPU-version (Notebookcheck sometimes measures CPU efficiency with CB R24) but thanks for the GPU info!
Edit: 46% for CB R24 GPU is impressive, but I do wonder, as has been brought up as a possibility several times in reviews, if the 60-70% we're seeing in Blender GPU, also a 3D renderer, is due to Apple leveraging the GPU neural accelerators there? I'm pretty sure in Blender Nvidia has OptiX code to do denoising using their tensor cores and I think Apple has worked hard to ensure parity, so maybe it was using the Neural Engine for that purpose and that code is now using the GPU? Because I'm seeing the same 60-70% increase on Blender 4.2 released in the summer so this isn't something new they just added and Metal 4 doesn't seem like it's really ready yet anyway. Blender was already Apple's best GPU renderer, so I'm thinking maybe it was already using the NE somehow and now the GPU matrix accelerators have just improved on that?
Notebookcheck shows a 44% uplift for CP2077 (no upscaling, or rt). 1080p Ultra.
View attachment 2570984
Test Apple MacBook Pro M5 (2025) - Der schnellste Single-Core-Prozessor der Welt
Notebookcheck testet das Apple MacBook Pro 14 M5 mit dem neuen M5-SoC, 32 GB RAM und Nanotexturglas-Mini-LED.www.notebookcheck.com
For what it's worth arstechnica claims to have tested the Mac version of CP 2077 and gotten basically the exact same numbers:
MacBook Pro review: Apple’s most awkward laptop is the first to show off Apple M5
Apple M5 trades blows with Pro and Max chips from older generations.
arstechnica.com
Scores are a little lower for the M4 but those were Airs, so that part is explainable.
But overall it doesn't make sense. I thought the port of CP2077 (especially once the settings bug was fixed) was much more performant than its emulated sibling?
Edit: yeah the native version should be 10-20% faster (and not sure if they had manually fixed the setting bug)
Cyberpunk 2077 - Crossover Windows vs MacOS Native Version - Mac gaming - Games on Mac- Macbook Games - Blendlogic
Cyberpunk 2077 on M1 Max: The Ultimate Mac Gaming Benchmark Welcome back, Mac gaming enthusiasts! Today, we’re taking a closer look at how one of the most demanding MacBook games, Cyberpunk 2077, performs on an M1 Max-powered MacBook Pro. After some great feedback from the community, I’m diving...
www.blendlogic.com
Cyberpunk 2077 has been tested on several Apple Silicon generations, results range from around 13 to 105fps at 1080p, depending on Mac and game settings tested
The Max chips, with their beefier GPUs, will serve you best in Night City.
Last edited:
People seem to forget that the Steam survey shows its Mac market share in percentage, not actual numbers. Even though it can seem that the number of Mac gamers hasn't grown the number of monthly active users on Steam grows by millions each year. This year Steam has 154 million active users and 147 million monthly active users. 1.91% of that is 2.81 million monthly active Mac users or 2.94 million active Mac accounts. So even if the Mac market share doesn’t grow the number of Mac users grows each year.
Steam Statistics 2025: Users, Revenue, Top Games & Market Trends
Steam Statistics 2025: Users, Revenue, Top Games & Market Trends
sqmagazine.co.uk
“To put the GPU to the test, I had to try out Cyberpunk 2077, which came out earlier this year on Mac. You can, to my surprise, play the game at 1920 x 1200 without upscaling and get decent frame rates at the Medium or Low graphics settings. Averaging 39 frames per second on Medium or 51 fps on Low is decent”
“I also spun up Lies of P and Resident Evil: Village, and saw similar results. You might have to reduce settings, but you can play these games natively now at playable frame rates, which is a big step forward.”
From Wired
“I also spun up Lies of P and Resident Evil: Village, and saw similar results. You might have to reduce settings, but you can play these games natively now at playable frame rates, which is a big step forward.”
From Wired
Doesn't Blender use Open Image Denoise? I don't recall that library using CoreML, and it certainly does not seem to use the new Metal tensor APIs. Not to mention that the denoising step shouldn't be dominant in these tests.
IMO the Blender results make perfect sense to me given what we know about M5 GPU. It's a combination of improved efficiency (new dynamic caching, new RT work scheduling hints etc.), improved compute (double half-precision), and faster memory.
No idea - had actually meant to tag you to get your input specifically
Just trying to figure out if there was a relationship between Blender already being the best GPU renderer for Apple and the even higher uplift in the M5 GPU other than the obvious "Apple engineers helped contribute to the open source code project". Because a 60-70% performance increase in a single generation is nuts, especially when a lot of other software, including other 3D renderers like CB R24 GPU, is seeing 28-45% increases (which by itself would be excellent).
Edit: did some digging with your suggestion and actually it's a maybe?
https://www.openimagedenoise.org/
v1.3.0
BNNS is part of the accelerate framework to do neural networks for training and inference
If it's not a bottleneck then maybe not the cause of the underlying speedup of Blender vs others, but an interesting thought. EDIT2: If Google's AI is to be believed most renderers including the one underlying Cinebench at least have OIDN as an option, so maybe not (however maybe it isn't the default one so maybe CB R24 GPU doesn't use it and whatever they do use isn't accelerated on the M5? - but the default is possibly OIDN).
Last edited:
If Steam's overall user base continues to grow, even if the Mac share stays at 2%, we could see 4–5 million Mac gamers in a few years. The M-series chips already made the Mac platform more appealing, and this is just the start.People seem to forget that the Steam survey shows its Mac market share in percentage, not actual numbers. Even though it can seem that the number of Mac gamers hasn't grown the number of monthly active users on Steam grows by millions each year. This year Steam has 154 million active users and 147 million monthly active users. 1.91% of that is 2.81 million monthly active Mac users or 2.94 million active Mac accounts. So even if the Mac market share doesn’t grow the number of Mac users grows each year.
Steam Statistics 2025: Users, Revenue, Top Games & Market Trends
Steam Statistics 2025: Users, Revenue, Top Games & Market Trends
My guess would be that Apple's involvement is likely a big factor. I wouldn't be surprised if Blender is more strategic about using half-precision, which would mean big boosts on Apple architectures specifically (while doing little for other vendors).
Also, dynamic caching is likely a very important part of the puzzle. Blender shaders are complex, with many dynamic paths, so occupancy really becomes a problem. I remember for M1/M2 Apple was submitting these patches where they would experimentally figure out the most performant dispatch size for a given model. For M3, all of this was replaced by a generic grid size with a comment "dynamic caching will sort it out". So improvements here will definitely have a large impact, especially in combination with RT improvements (looking at new RT-related patents, it's all about improving occupancy and reducing synchronization overhead). I am also certain that Apple engineers broke down the shaders so that the occupancy is maximized on their hardware. I doubt that Redshift is as mature.
Finally, nature of the benchmarks also probably plays a role. CB R24 is essentially a tech demo. It uses a single scene, which is not even that complex. Blender benchmark uses multiple scenes, so there is higher chance to hit cases where M5 performs particularly well.
That's probably accurate however I think there might be a little something to this idea of denoising being improved on the M5. While you were typing out your response, I was researching your mention of OIDN and editing my post above.
OIDN is using Apple's accelerate framework (BNNS to be exact) and can use Apple GPUs, so it's possible the M5 neural cores are accelerating denoising if BNNS/Accelerate has been updated to do so. Now Redshift can also use OIDN for denoising but it also has two other options for denoising (including OptiX) and so it isn't clear to me if CB R24 is doing so. Obviously if AI-accelerated denoising was giving Blender that extra boost, then CB R24 should be getting it too if it also uses OIDN, but not if doesn't. A lot of ifs here so not going hard on this theory, but it was an interesting rabbit hole.
Aye, we already know that not every game engine and renderer are being boosted the same amount - some as "low" as 28%, most around ~40%, Blender being about 60%-70% on the three scenes chosen.
To make it perfectly clear — all of this is speculation. We won't know what contributes most to these improvements without carefully profiling the shaders and comparing the data across models. Which would be a lot of work, and frankly, probably not the best use of one's time.
The performance boost will always be workload-dependent. It would be much simpler to reason about if the GPU core remained the same and only the clocks + memory interfaces were boosted. But we are dealing with complex changes to the GPU core that will interact with the software in a way that can be difficult to predict. Overall, I'd speculate that the improvements will be smallest on compute-heavy workloads that use FP32 precision and have predictable resource footprint, and it will be largest on shaders that use half-precision aggressively, as well as shaders that rely on complex asymmetric control flow. RT is a big thing as well, since we don't actually know what Apple did to improve it. As I mentioned, there are some new RT-related patents that focus on efficiency and caching, which should interact well with the dynamic caching feature, but they might have also increased the number of RT processors or similar.
With as many macs out there as there are the absolute number of mac gamers using steam isn't outpacing the growth of the the steam userbase. It does/can cause questions like: Why has Capcom ported the RE entries, but none of the Monster Hunter ones that use the same engine? Where is Death Stranding 2? Etc.People seem to forget that the Steam survey shows its Mac market share in percentage, not actual numbers. Even though it can seem that the number of Mac gamers hasn't grown the number of monthly active users on Steam grows by millions each year. This year Steam has 154 million active users and 147 million monthly active users. 1.91% of that is 2.81 million monthly active Mac users or 2.94 million active Mac accounts. So even if the Mac market share doesn’t grow the number of Mac users grows each year.
Steam Statistics 2025: Users, Revenue, Top Games & Market Trends
Steam Statistics 2025: Users, Revenue, Top Games & Market Trends
1200p with no upscaling seems like an odd setup. Surely 1080p with MetalFX at upscaling from 720p or 900p would be the way most people would play this?