Question AMD 890M vs Apple M3

[mikegg]

4k Aztec High GFX

* AMD 890M: 39.1fps
* M3: 51.8fps

3DMark Wild Life Extreme

* AMD 890M: 7623
* M3: 8286

Power:

* AMD 890M: 46w
* M3: 8286: 17w

M3 about ~253% more efficient in these benchmarks.

Transistor count:

* AMD Strix Point: Estimated 40 billion based on die size and N4P
* M3: 25 billion

Source: https://www.notebookcheck.net/AMD-Z...le-M3-and-Qualcomm-Adreno-X1-85.868475.0.html

 

[gdansk]

Wow that's crazy it appears ARM is simply better than x64 in every way by a huge margin

That puts Apple ahead of Nvidia on efficiency too. They should stop making chips and pay Apple to do it for them

Edit: In Baldurs Gate 3 it seems the 890M is almost 40-50% faster than the M3. So maybe it doesn't actually translate to games.

 

[inquiss]

Wow that's crazy it appears ARM is simply better than x64 in every way by a huge margin

That puts Apple ahead of Nvidia on efficiency too. They should stop making chips and pay Apple to do it for them

Edit: In Baldurs Gate 3 it seems the 890M is almost 40-50% faster than the M3. So maybe it doesn't actually translate to games.

Yeah it doesn't translate to games at all. Synthetic ocmparisons are pointless for Apple silicon.

 

[Mopetar]

Yeah it doesn't translate to games at all. Synthetic ocmparisons are pointless for Apple silicon.

How much of that is just an indicator of unpolished Apple drivers though? They don't really have a history of caring about games, but the synthetic benchmarks indicate that the hardware has potential.

I wouldn't mind at all if Apple hired some engineers to work on improving their GPU drivers for games. Even if they just focus on a small number of the most popular games I think it would go a long way towards more companies supporting Macs in the future.

 

[inquiss]

How much of that is just an indicator of unpolished Apple drivers though? They don't really have a history of caring about games, but the synthetic benchmarks indicate that the hardware has potential.

I wouldn't mind at all if Apple hired some engineers to work on improving their GPU drivers for games. Even if they just focus on a small number of the most popular games I think it would go a long way towards more companies supporting Macs in the future.

Yeah, maybe. But they haven't? So it's like not relevant in games which is what the comparison should be, no one plays timespy you know?

 

[CouncilorIrissa]

Highest effort senttoschool bait lmao

 

[poke01]

Apple GPUs are good but they suck for gaming cause macOS, no Vulkan, no Direct X support. However, in non-gaming tasks they have the best iGPUs, bar none. Plus Mac games are not optimised at all.

but for Blender, video editing and 3D rendering they are excellent for a laptop and for the power they consume.
890M:

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

M3:

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

Cinebench 2024 GPU:
M3 - ~3340
890M - ~1157

 

[poke01]

Highest effort senttoschool bait lmao

I mean the Apple GPUs beat the crap out of AMD GPUs in non-gaming tasks. That reddit thread is crazy cope though.

I think people don't yet reliase how good the M3 GPU redesign was...

 

[poke01]

Also the M3/M4 was Apple's first take at RT and its better than Intel's (yes even Lunars) and AMD's. Its not talked about much because most talk about GPUs is gaming but if one watchs the WWDC talks about the GPUs its pretty interesting. Apple implemented some things that not even Nvidia did yet. But don't expect great gaming performance this is strictly for non-gaming.

 

[DavidC1]

4k Aztec High GFX

* AMD 890M: 39.1fps
* M3: 51.8fps

I think many of the mobile compatible benchmarks such as Aztec Ruins runs FP16 on mobile hardware while it does FP32 on PCs. That'll be responsible for greater than 20% difference, since it doubles compute capability.

Time Spy is an OK cross platform benchmark in this case. You'll see Snapdragon loses significantly in TS while it's about equal on others. If comparing against mobile chips, you should suspect FP16 vs FP32.

Time Spy is also quite demanding on iGPUs, so it's representative of games that run at High settings, which most won't do because it's hard to reach 60 fps stable. Thus for iGPUs it overestimates performance differences. In case of ARC, the architectural deficiency of needing high workload to maximize utilization hurts it in games, since it runs at lower settings unlike Time Spy.

You can see that Cyberpunk gains are similar to Time Spy. In Strange Brigade, the gains go from low in Low settings to higher in High settings.

GPUs are harder to differentiate and mostly depend on Moore's Law to advance. That's because embarassingly parallel workloads benefit from just having more compute.

CPUs are harder because it actually requires smarts - execution, innovation, and creativity from engineers.

How much of that is just an indicator of unpolished Apple drivers though? They don't really have a history of caring about games, but the synthetic benchmarks indicate that the hardware has potential.

Doubt it's drivers as the hardware lacking in more ways than it shows itself.

ARC drivers were solely blamed for the same thing. Now we know a significant amount of problems are due to the hardware.

Suspect mobile benchmarks. When Voodoo came out, it was revolutionary because they realized in consumer graphics, it didn't need the IEEE FP64 precision, and they could get away with much less accurate FP32.

In mobile, it further exaggerates the fact because smaller screen, shorter play and much less accurate targeting lowers the need for detail even further. Thus FP16 works. There are likely other details it sacrifices.

 

[poke01]

I think many of the mobile compatible benchmarks such as Aztec Ruins runs FP16 on mobile hardware while it does FP32. That'll be responsible for greater than 20% difference, since it doubles compute capability.

GFX Bench is a horrible benchmark. I only consider 3D Mark TS or Steel Nomad for synthetics.

In case of ARC, the architectural deficiency of needing high workload to maximize utilization hurts it in games, since it runs at lower settings unlike Time Spy.

This is something Appe took into consideration in the M3 GPU design. https://forum.beyond3d.com/threads/apple-dynamic-caching-on-m3-gpu.63419/

 

[DavidC1]

GFX Bench is a horrible benchmark. I only consider 3D Mark TS or Steel Nomad for synthetics.

TS is actually a decent test. I would call it SpecCPU of GPU tests.

I look for one thing in a sythetic test, how it performs in actual games. When people are running synthetic tests, they want it comparable, so it runs the default settings. It runs at 1440p and what's considered Highest/Ultra settings.

This makes sense for dGPUs that go in a PCI Express slot, but is very demanding, actually not too practical for iGPUs.

Actual games have to bring the settings and resolution low to make it practical. Yes you can do the same with Time Spy using lower settings but the Default settings is what most people use, and makes comparison harder.

Now every generation changes the ideal benchmark. It used to be 3DMark06, then 11, then I forgot about it for a while, but it's now Time Spy.

This is something Appe took into consideration in the M3 GPU design. https://forum.beyond3d.com/threads/apple-dynamic-caching-on-m3-gpu.63419/

Yet the performance in gaming sucks. Comparison with other ARM vendors show their GPUs aren't that good comparatively as mobile benches indicate.

Drivers can make a difference between one winning and not winning but when the differences between two parts in the order of nearly 2x, then there's something else going on.

Don't simply disregard results. Think of why it looks strange.

 

[poke01]

Yet the performance in gaming sucks. Comparison with other ARM vendors show their GPUs aren't that good comparatively as mobile benches indicate.

Drivers can make a difference between one winning and not winning but when the differences between two parts in the order of nearly 2x, then there's something else going on.

It sucks because these GPUs while good are not a priority for game devs because of low marketshare and Apple being hostile to games. Apple also gives no crap about gaming unlike Sony or MS. They don’t even have a specific gaming division. Apples funding is all towards creative tools/apps ie blender support, Maxon, Blackmagic etc.

 

[DavidC1]

It sucks because these GPUs while good are not a priority for game devs because of low marketshare and Apple being hostile to games.

Yes, some of that is true but I'm saying a significant portion is likely due to the compromises they make that doesn't really matter for mobile, such as sacrificing details(FP16 vs FP32). Literally the mobile chips are running FP16 to get their scores and the PC parts are using FP32. That's 2x the compute(Flops) advantage for mobile!

How else would you explain it sucking in Time Spy but being stellar in mobile benchmarks? Simply the mobile benchmarks do not reflect PC gaming performance at all.

Back when Intel had mobile Atom parts, their graphics was doing fantastically in mobile graphics benchmarks relative to their own, even though it used the same GPU architecture.

 

[Mopetar]

Doubt it's drivers as the hardware lacking in more ways than it shows itself.

It wouldn't do as well as it does in synthetics or other applications if that were the case.

I suppose you could argue that the Apple GPUs are missing some necessary component that holds gaming back, but I can't think of any reason why Apple would gimp gaming performance over what's likely to be a trivial number of transistors.

The more simple explanation is that the game developers aren't going to bother optimizing for Metal on top of DX12/Vulkan and most Mac users are happy enough for whatever gaming scraps they can get that none of them really care if the performance isn't top notch.

 

[DavidC1]

It wouldn't do as well as it does in synthetics or other applications if that were the case.

I already said one reason why it does well in mobile benchmarks and does terrible in desktop ones. FP16 vs FP32. The mobile vendors do a good job in GPUs, but it's nowhere as good as mobile benches say they are.

 

[LightningZ71]

You have to keep in mind that there is a foundational difference between iGPUs like the AMD 890M and the M3: available effective memory bandwidth. The M3 has a very well executed SLC of usable size that gives it an EFFECTIVELY much higher apparent VRAM bandwidth as opposed to the 890M, which tops out at about 120GB/s with LPDDR5X-7500. That makes a big difference in both iGPU performance as well as iGPU power consumption. Accesses that can be served by the SLC are much less expensive in energy usage. In addition, for many parts of ray tracing, they thrash small sections of VRAM excessively. Having a superfast SLC can drastically speed up those functions.

This isn't to excuse AMD as they could have chosen to put a MALL cache on Strix Point (and even planned to do so at one point, until MS started pushing copilot and forced them to have a large NPU to be feature competitive). They didn't and here are the results.

 

[misuspita]

And a NPU that is useless...

 

[LightningZ71]

Useless TODAY, possibly forever. IMHO, they should have stuck with a much smaller one, done a MALL cache, and tweaked RDNA3.5 a bit to improve it's AI performance enough to qualify the platform. But that's just me..

 

[DZero]

Makes me think what will happen if Apple decides to add at least 8 GB of GDDR6 on their GPU as dedicated memory. It would change even more the game? Would be on par of mid or high tier dGPU offerings?

 

[misuspita]

Useless TODAY, possibly forever. IMHO, they should have stuck with a much smaller one, done a MALL cache, and tweaked RDNA3.5 a bit to improve it's AI performance enough to qualify the platform. But that's just me..

They were as usual [@#$%^] by Microsoft, not integrating it in the Windows functionality. I would never buy a APu fr the NPU functions, or AI. It's useless for now, when I type this, and for the life of the product. And yes, MALL on the size of the NPU would have done much better but I am a couch CPU designer so....

 

[Mopetar]

Makes me think what will happen if Apple decides to add at least 8 GB of GDDR6 on their GPU as dedicated memory. It would change even more the game? Would be on par of mid or high tier dGPU offerings?

I don't think they need to do that. They're already using soldered memory that clocks faster than standard DDR RAM sticks and they have a fairly wide bus that already gives them more effective bandwidth than many low-end GPUs. Some of that has to be shared with the CPU, but they also have a large SLC that the GPU can use as well.

I think the main thing holding them back are drivers or just a lack of interest in embracing open standards like Vulkan instead of pushing their own Metal API. Developers aren't going to want to spent the extra time to optimize for something that's only used by a tiny part of the overall market. Mac users are typically happy enough just to get a port that will run on their hardware at all, even if it's not the best in terms of performance.