They are not so good

[99 others]

I just saw this video on youtube from Linus Tech tips and they tested the M1 Max :

Not so good after all compared to a lot of x86 systems out there, at least graphic wise.
What do you think?
Still waiting for my MBP M1 PRO 16…

 

[mcnallym]

If wanted to run the tasks where the x86 does better I would not be buying an Apple anyway.

Rosettes, OpenGL and Vulcan apps not going to show any M1 in a good light compared to x86 machines.

As they say it shines where the Apps written to use the metal api, and can use the hardware, ie video editing.

Outside of that the the max going to be overkill.

Waiting for a mini pro/max myself but do use fcp x and video editing so will work out for me.

X86 and windows not going to be an option for me.

 

[thedocbwarren]

This is such a a tired argument. Just buy an bloody Intel if you want to game on Windows. Intel is a dead-end on mac regardless of x86 or not. Given time things change up but if Intel and gaming is the need then a solution exists. Sorry it's one thread after another... over and over again.

 

[l0stl0rd]

I just saw this video on youtube from Linus Tech tips and they tested the M1 Max :

Not so good after all compared to a lot of x86 systems out there, at least graphic wise.
What do you think?
Still waiting for my MBP M1 PRO 16…

Ther are fine… 14” is slightly “underclocked” also shame nobody compares to AMD…

 

[ahurst]

Like Anthony says in the video, I wonder how much of this is due to poor optimization for the new architecture/metal and how much is due to their fundamental GPU design. In the Anandtech review there were synthetic GPU benchmarks where the M1 Max did really well (coming close to the mobile 3080), so I guess time will tell if those were anomalies.

In practical terms *today* that question doesn't really matter: if it's slower for the job in practice (e.g. that 3D renderer discussed in the video), then it's going slower, regardless of the reasons why. The main hope is that with more time and developer adoption, there'll be some healthy room for future improvement with further optimization.

Maybe this is silly, but if I were a clever Apple executive trying to spur better optimization for the new GPUs, I'd slyly send out free units to some top emulator developers. It's an area where code is generally open-source (so other people can learn from the examples and optimizations), where people try their best to milk out good performance from the hardware, and where Metal has a near-zero adoption rate. At the very least, I think you'd see major improvements to MoltenVK and maybe a modern OpenGL 4.6 to Metal translation layer out of it for minimal up-front investment.

 

[uller6]

I just saw this video on youtube from Linus Tech tips and they tested the M1 Max :

Not so good after all compared to a lot of x86 systems out there, at least graphic wise.
What do you think?
Still waiting for my MBP M1 PRO 16…

Is this paid astroturfing from Linus tech tips? It would be more helpful if you summarized the video in words instead of adding the link as clickbait.

 

[Kronekerdelta]

This is such a a tired argument. Just buy an bloody Intel if you want to game on Windows. Intel is a dead-end on mac regardless of x86 or not. Given time things change up but if Intel and gaming is the need then a solution exists. Sorry it's one thread after another... over and over again.

The benchmarks are lackluster in many applications besides gaming.

 

[thedocbwarren]

The benchmarks are lackluster in many applications besides gaming.

uh yeah, no. Sorry, not true. Deny all you like, Apple Silicon is highly performant plus low power/heat. It's the largest gain in performance vs anything on the PC side. Again, if anyone is hell bent on Intel and gaming, or whatever nonsense these threads feel like comparing for the sole sake of bashing, then by all means buy a PC. In the real world the gains are obvious.

 

[UBS28]

I just saw this video on youtube from Linus Tech tips and they tested the M1 Max :

Not so good after all compared to a lot of x86 systems out there, at least graphic wise.
What do you think?
Still waiting for my MBP M1 PRO 16…

Benchmarks don't tell everything. The 16" M1 Max stays dead quiet when used the maximum and the battery life is far superior.

Also the M1 Max is a waste on the 14" MBP, only the 16" MBP can handle it.

As a laptop, the 16" M1 Max MBP is the best on market right now. For a gaming laptop, hell no. These thing suck for gaming because of no Bootcamp support anymore.

 

[l0stl0rd]

Benchmarks don't tell everything. The 16" M1 Max stays dead quiet when used the maximum and the battery life is far superior.

Also the M1 Max is a waste on the 14" MBP, only the 16" MBP can handle it.

As a laptop, the 16" M1 Max MBP is the best on market right now. For a gaming laptop, hell no. These thing suck for gaming because of no Bootcamp support anymore.

Yup silence is great, I did a render in the self compiled Blender 3.1 and took 6min and the fans did not even turn on (on the M1 Pro).

Even rendering on both the GPU and CPU at the sane time does like 2.2 K rpms that nothing, depending what room you are in the background noise is louder

And I agree 14” and Max is not a choice I would make.

 

[jeanlain]

Let's stop finding excuses. Apple said the M1 Max can compete with the 3080 mobile on performance (not just perf/W), which is very rarely the case.

Forget about compute. The M1 Max has 10 TFLOPs vs 19 TFLOPs for the 3080 mobile. It cannot mathematically compete, as tests confirm (disappointing results in geekbench Metal, tensor flow, ect.).
Forget about ray tracing, the M1 doesn't have dedicated hardware. It's handily beaten by RTX GPUs.
Forget about gaming. There's only one Mac game truly optimised for Apple Silicon. In a couple of years, we may have another if we're lucky.
That leaves what? FCPX, Davinci resolve + some iOS-derived benchmarks like GFXBench?

Yeah the 3080 is louder, hotter, whatever. But Apple said the M1 was almost as fast, while it clearly is not. They also said the M1 Max was 3 times faster on battery, a result that non one could reproduce.
Apple set expectations way too high.

 

[leman]

Like Anthony says in the video, I wonder how much of this is due to poor optimization for the new architecture/metal and how much is due to their fundamental GPU design.

The GPU design is fine. It is obviously not as mature as Nvidia, but it's more streamlined and has more cache to work with, so it ends up much more energy efficient for many things. It is a different matter that we don't have much software that can properly take advantage of it.

Let's stop flinding excuses. Apple said the M1 Max can compete with the 3080 mobile on performance (not just perf/W), which is very rarely the case.

I agree that for general-purpose compute there is no competition, unless we take some more contrived cases of complex computations where M1 is likely to pull ahead because of its humongous cache. M1 will also probably be better where you are working with very large data (it has much more RAM available to it).

M1 Max rasterization performance should be close enough to a mobile 3080 — so in a well optimized game it should do just as well (without RT of course, that is again absolutely no competition).

Finally, M1 Max will do better on many productivity workloads simply because of lower data shuffling overhead.

 

[jeanlain]

I'd still like to see the task where the M1 Max is 3x faster than the RTX 3080 on battery, like Apple said. Even in the most favourable case (GFXBench) the M1 Max is far from being 3x faster.

 

[crazy dave]

Let's stop flinding excuses. Apple said the M1 Max can compete with the 3080 mobile on performance (not just perf/W), which is very rarely the case.

Forget about compute. The M1 Max has 10 TFLOPs vs 19 TFLOPs for the 3080 mobile. I cannot mathematically compete.
Forget about ray tracing, the M1 doesn't have dedicated hardware. It's handily beaten by RTX GPUs.
Forget about gaming. There's only one Mac game truly optimised for Apple Silicon. In a couple of years, we may have another if we're lucky.
That leaves what? FCPX, Davinci resolve + some iOS-derived benchmarks like GFXBench?

Yeah the 3080 is louder, hotter, whatever. But Apple said the M1 was almost as fast, while it clearly is not. They also said the M1 Max was 3 times faster on battery, a results that non one could reproduce.
Apple set expectations way too high.

I can’t remember if they said it explicitly but even if they didn’t, it’s pretty clear that Apple was referring to GFXBench Aztec High. And you can replicate their results there. The reason it performs so well is it is a graphics rasterization workload that doesn’t use ray tracing and has long been cross platform being built to benchmark phones as well as computers. So it was already optimized for both TBDR and IMR GPUs.

But yes the issue is that it takes time for software to get optimized. Most software released before the M1 is launched won’t be and performance will be underwhelming.

Redshift and blender are both saying further optimizations and performance improvements are coming, and gaming is going to take longer. So yeah I think you can blame Apple for overhyping the promise but not acknowledging the time that it’s going to take for the software to actually practically mature.

And yes in terms of raw TFLops for 32bit compute a 10TFLOP GPU simply won’t compete with a mobile 3080.

 

[jeanlain]

I can’t remember if they said it explicitly but even if they didn’t, it’s pretty clear that Apple was referring to GFXBench Aztec High.

No Apple did not disclose the tests they used.
Maybe it was a combinaison of GFXBench Aztec (where the M1 Max is actually faster than the competition), 3DMark Wild Life, and possibly Basemark GPU (an obscure test that happens to run well on Apple Silicon).

 

[jeanlain]

Redshift and blender are both saying further optimizations and performance improvements are comin

Yes, but the M1 Max will still be far behind given its lower FLOPs and absence of RT hardware. Perhaps it can compete on scenes taking a lot of memory.

 

[leman]

And yes in terms of raw TFLops for 32bit compute a 10TFLOP GPU simply won’t compete with a mobile 3080.

A quick note about this: the TFLOPS specs given by manufacturers are highly misleading. They are essentially the amount of basic floating point operations (additions and multiplications) that a GPU can perform per second when running a sequence of fused-multiply-add instructions and nothing else.

Modern NVIDIA GPUs have two sets of ALUs that can work in parallel: one for FP computation and one for INT computation. With Ampere, Nvidia gave the INT ALU the capability to execute FP32 instructions as well, which basically doubled the FP32 TFLOPS figures. Of course, you don't see these kind of improvements on normal workloads since shaders do use INT operations and other complex tuff, and they do use data loads and stores, which add latency and yes things up in general. So take this XXX TFLOPS with a big grain of salt.

Of course, same applies to Apple Silicon which does not have dedicated INT ALUs — it's ALU can do either INT or FP operation per cycle. So Nvidia has a clear lead here with its dual ALU architecture.

 

[gigatoaster]

This is a very interesting conversation but in the end what is the point of all is? For you guys, what does it mean that this chip is better or worst than the other one?

Why looking at specs and not use-cases?

 

[crazy dave]

I'd still like to see the task where the M1 Max is 3x faster than the RTX 3080 on battery, like Apple said. Even in the most favourable case (GFXBench) the M1 Max is far from being 3x faster.

Looking at their numbers it’s definitely GFXBench they were quoting. Okay I don’t know about definitely but it’s the closest. The claim is that it’s 2.5x faster and I think it’s a Razer that they’re comparing against but the version I’m looking at is fuzzy and I can’t read it right. Should be possible to verify.

 

[leman]

This is a very interesting conversation but in the end what is the point of all is? For you guys, what does it mean that this chip is better or worst than the other one?

Why looking at specs and not use-cases?

Because many of us here are interested in technical details. It's not about "better or worse" per se. It's about learning how all this stuff works.

Personally, I find Apple's approach technologically more interesting than the traditional brute-force GPUs built by Nvidia and AMD. There is certain elegance to TBDR and streamlined GPU ALUs that others lack. Of course, Nvidia has superior technology in many domains (they have been at it for a while after all), but Apple is catching up quickly. Once they have hardware RT and maybe more capable ALUs, their hardware will be top tier. Traditional GPUs on the other hand have no space to grow — as one can clearly see from the recent history. Nvidia's main performance improvements came from increasing the power consumption of their GPUs.

 

[crazy dave]

A quick note about this: the TFLOPS specs given by manufacturers are highly misleading. They are essentially the amount of basic floating point operations (additions and multiplications) that a GPU can perform per second when running a sequence of fused-multiply-add instructions and nothing else.

Modern NVIDIA GPUs have two sets of ALUs that can work in parallel: one for FP computation and one for INT computation. With Ampere, Nvidia gave the INT ALU the capability to execute FP32 instructions as well, which basically doubled the FP32 TFLOPS figures. Of course, you don't see these kind of improvements on normal workloads since shaders do use INT operations and other complex tuff, and they do use data loads and stores, which add latency and yes things up in general. So take this XXX TFLOPS with a big grain of salt.

Of course, same applies to Apple Silicon which does not have dedicated INT ALUs — it's ALU can do either INT or FP operation per cycle. So Nvidia has a clear lead here with its dual ALU architecture.

That’s true that’s a nuance I was glossing over - another factor is that neither are great for FP64 performance but I’m not sure if anyone has even measured that on the M1.

 

[crazy dave]

Yes, but the M1 Max will still be far behind given its lower FLOPs and absence of RT hardware. Perhaps it can compete on scenes taking a lot of memory.

I think the FLOPs will matter less than the lack of RT hardware. So hopefully Apple will add that soon as that will make a big difference. Was a bit disappointed that the A15 GPU didn’t have it. Hopefully the A16/M3 will. Maybe tensor cores too while we’re at it?

 

[crazy dave]

Traditional GPUs on the other hand have no space to grow — as one can clearly see from the recent history. Nvidia's main performance improvements came from increasing the power consumption of their GPUs.

To be fair the more traditional designs have advantages with respect to anti aliasing and DLSS techniques. So I wouldn’t necessarily claim it’s a clean intrinsic advantage for TBDR. It’s a trade off.

 

[leman]

That’s true that’s a nuance I was glossing over - another factor is that neither are great for FP64 performance but I’m not sure if anyone has even measured that on the M1.

Oh, that’s simple - Apple GPUs do not have any hardware support for FP64. Which makes perfect sense to me: given how crippled GPU FP64 usually is, one is better off implanting extended precision in software. Predictable performance and behavior.

To be fair the more traditional designs have advantages with respect to anti aliasing and DLSS techniques. So I wouldn’t necessarily claim it’s a clean intrinsic advantage for TBDR. It’s a trade off.

I dint see why they would? AA is essentially “free” on TBDR (it’s all on tile, you don’t have to do expensive sample memory fetches) and for DLSS there is no difference. Actually, TBDR might have a slight advantage here too because of more efficient memory transfers.

Where traditional GPUs have an advantage is in techniques like mesh shading which are not a good fit for a TBDR.

 

[crazy dave]

Oh, that’s simple - Apple GPUs do not have any hardware support for FP64. Which makes perfect sense to me: given how crippled GPU FP64 usually is, one is better off implanting extended precision in software. Predictable performance and behavior.




I dint see why they would? AA is essentially “free” on TBDR (it’s all on tile, you don’t have to do expensive sample memory fetches) and for DLSS there is no difference. Actually, TBDR might have a slight advantage here too because of more efficient memory transfers.

Where traditional GPUs have an advantage is in techniques like mesh shading which are not a good fit for a TBDR.

Do you know what the software FP64 performance of M1 is? I’m just curious.

How I remember it being explained was that DLSS and any AA technique that required information on the whole scene/larger-than-a-tile area effectively require a second rendering pass for TBDR GPUs. I’ll see if I can dig up the link.

Edit: having trouble finding it quickly and it’s late here so may try again later.