So M1 Max with 32 Core GPU matches RTX 3080 Mobile.

[crazy dave]

Also, from what I understand, isn't Metal a subsection or something to Vulkan? From what I understand, Vulkan runs pretty damn well on macOS.

No Vulkan is not supported on AS macOS, AMD GPU Macs might - but generally when used it gets translated into metal by moltenvk (open source not an Apple product, must be used by the program itself).

Metal is a different API with some major some minor differences between it and Vulkan. @leman can go into more detail. I’m a CUDA guy, but he’s more familiar with those.

Edit: And he already responded ?

 

[Pressure]

Most of the benchmarks are probably running plugged in.

I’m sure things get interesting once they are run on battery only.

 

[ElfinHilon]

No, Metal and Vulkan are completely different and have nothing in common.



There is an open source project that implements Vulkan on top of Apple Metal. It works fine.

Ah got ya. How good is Vulkan over Metal?

 

[mi7chy]

Also, from what I understand, isn't Metal a subsection or something to Vulkan? From what I understand, Vulkan runs pretty damn well on macOS.

Vulkan was built on Mantle.

 

[PlainBelliedSneetch]

Nope, the 512bit memory bus (running @ 400GB/s) is on the M1 Max SoC, which you can get with either 32GB or 64 GB of RAM...

So the 32GB still has 4 memory chips to get the extra bandwidth.

 

[jmho]

Ah got ya. How good is Vulkan over Metal?

Metal will always be better than Vulkan on Macs, because Metal is designed specifically for Apple hardware while Vulkan needs to support far more cards from multiple vendors all with different architectures. Vulkan is still very nice though.

Personally I think the API question is a bit of a red-herring. Most games consoles have had some form of support for OpenGL but nobody uses OpenGL on consoles because it's a massive waste of performance. Most developers would rather write their own abstraction layer that calls the fastest API available on each platform.

I looked at Apple's recent pull request to Blender where they added a bunch of very very minor tweaks to the shaders which already compile for OpenCL and OptiX to make them compile for Metal and it was not a huge job at all. It just needed someone with the skills and motivation to actually do it.

I think motivation has been the major thing holding back 3D on the Mac until now (hopefully), because what's the point in spending time on the software if the hardware hasn't historically been up to the task?

 

[ElfinHilon]

Metal will always be better than Vulkan on Macs, because Metal is designed specifically for Apple hardware while Vulkan needs to support far more cards from multiple vendors all with different architectures. Vulkan is still very nice though.

Personally I think the API question is a bit of a red-herring. Most games consoles have had some form of support for OpenGL but nobody uses OpenGL on consoles because it's a massive waste of performance. Most developers would rather write their own abstraction layer that calls the fastest API available on each platform.

I looked at Apple's recent pull request to Blender where they added a bunch of very very minor tweaks to the shaders which already compile for OpenCL and OptiX to make them compile for Metal and it was not a huge job at all. It just needed someone with the skills and motivation to actually do it.

I think motivation has been the major thing holding back 3D on the Mac until now (hopefully), because what's the point in spending time on the software if the hardware hasn't historically been up to the task?

When you say red-herring, what do you mean? Red-herring towards which question?

 

[jmho]

When you say red-herring, what do you mean?

It's something that people get worked up about, but isn't really that important.

In a modern 3D engine the actual API layer is incredibly thin.

 

[ElfinHilon]

So I've stated this before, but I didn't provide any evidence.

Here's the 16core GPU compared to M1:
MacBookPro18,3 (16-Core GPU) vs Mac mini (Late 2020)

Here's the supposed 32core GPU compared to M1:
MacBookPro18,2 ("32"-core GPU) vs Mac mini (Late 2020)

Seeing these results, I do believe that Geekbench is reporting this as an error in the GPU core, when the reality is that the GPU being tested is indeed the 24core GPU.

Also, I reorganized these. I urge everyone to look at them. They look remarkably similar to what a theoretical 24 core GPU should have been giving us. While there is around a ~10% performance bump, this can easily be explained by having access to more RAM AND the greater bandwidth that comes with the Pro to Max.

 

[ElfinHilon]

It's something that people get worked up about, but isn't really that important.

In a modern 3D engine the actual API layer is incredibly thin.

And is this more in relation to the Metal vs Vulkan question I posted or more in relation to the Metal vs OpenCL benchmark scores we are seeing here?

 

[robertosh]

Do you think we will be seeing soon MAC MINIS or MAC PRO (DESKTOP) with the new chips as well?

I think so, they still need to replace the 64G intel option in the mini.

 

[jmho]

And is this more in relation to the Metal vs Vulkan question I posted or more in relation to the Metal vs OpenCL benchmark scores we are seeing here?

I'm just saying that Metal is the gold standard for finding out what the hardware can do.

For any benchmarks using Vulkan or OpenCL you don't know what's hardware and what's software / driver / compatibility overhead.

Obviously we don't have a lot of benchmarks yet, so the OpenCL benchmarks are interesting, but we have to take them with a grain of salt.

 

[ElfinHilon]

I'm just saying that Metal is the gold standard for finding out what the hardware can do.

For any benchmarks using Vulkan or OpenCL you don't know what's hardware and what's software / driver / compatibility overhead.

Obviously we don't have a lot of benchmarks yet, so the OpenCL benchmarks are interesting, but we have to take them with a grain of salt.

Right yeah, that's I've been saying. I'll be worried once we start seeing a number of other benchmarks showing similar-like performance to this.

 

[Irishman]

They told no such thing and the 32GB confit has the same bandwidth as the 64GB config. This is not about the amount of RAM but about the width of the RAM interface.

What I meant was that Apple made it no secret that "you have to get 64GB for multi-channel memory bandwidth otherwise it probably diminishes with every step down."

I'm not talking about the amount of RAM either.

 

[mi7chy]

Metal will always be better than Vulkan on Macs, because Metal is designed specifically for Apple hardware while Vulkan needs to support far more cards from multiple vendors all with different architectures. Vulkan is still very nice though.

Metal is a lock-in users/lock-out developers API. If it's that much better than Vulkan then offer both and let devs/users gravitate towards the better one. Same argument can be made that DirectX 12 is better since it's designed for Windows but reality is Vulkan is significantly better at maximizing hardware performance.

 

[jmho]

Metal is a lock-in users/lock-out developers API. If it's that much better than Vulkan then offer both and let devs/users gravitate towards the better one. Same argument can be made that DirectX 12 is better since it's designed for Windows but reality is Vulkan is significantly better for actual gameplay.

They do support both. MoltenVK is right there if you don't care about getting the best features / performance.

 

[mi7chy]

They do support both. MoltenVK is right there if you don't care about getting the best features / performance.

That's a Metal to Vulkan wrapper that puts penalty on Vulkan performance which reinforces Metal as a lock-out API.

 

[leman]

What I meant was that Apple made it no secret that "you have to get 64GB for multi-channel memory bandwidth otherwise it probably diminishes with every step down."

I'm not talking about the amount of RAM either.

Hm, its quite late here and I am not sure whether I understand your post correctly. So I‘ll just going to say what I know and if you mean the same thing then it’s great.

Anyway, 32GB and 64GB on Max have identical bandwidth of 400GB/s. The 32GB on Pro have bandwidth of 200GB/s due to different RAM topology.

 

[Irishman]

Userbenchmark is trash. Don't use em.

Why do you think so?

 

[ElfinHilon]

Hm, its quite late here and I am not sure whether I understand your post correctly. So I‘ll just going to say what I know and if you mean the same thing then it’s great.

Anyway, 32GB and 64GB on Max have identical bandwidth of 400GB/s. The 32GB on Pro have bandwidth of 200GB/s due to different RAM topology.

This is my understanding as well. I haven't heard anything saying that upgrading to 64GB of RAM gives you faster bandwidth.

 

[jmho]

That's a Metal to Vulkan wrapper that puts penalty on Vulkan performance which reinforces Metal as a lock-out API.

Metal is an API that is designed to sit as close to the "metal" as possible, and as such it creates everything in the perfect and optimal format to run at high speed on Apple hardware.

If you're writing Vulkan code and shaders that weren't specifically designed to run on Apple hardware I'm not sure how you intend to solve that conversion problem without some driver overhead or a translation layer?

It's like complaining that you have to speak English to an English speaker, when you really want to speak German. Apple GPUs speak metal. If you want to speak another language, you're going to need a translator.

 

[crazy dave]

Why do you think so?

Userbenchmark has been caught doing some questionable things where their reported results and especially their descriptive conclusions don’t align with anybody else’s and if I remember right with respect to Intel vs AMD CPUs, show blatant, obvious favoritism towards Intel. Like not not even trying to be subtle about it.

Most of the time when people complain about that sort of thing (tech outlet X or benchmark website Y is so biased!), I don’t find it very … compelling. But in a few of the links that were posted, even I had to admit, it was pretty bad.

 

[crazy dave]

Metal is an API that is designed to sit as close to the "metal" as possible, and as such it creates everything in the perfect and optimal format to run at high speed on Apple hardware.

If you're writing Vulkan code and shaders that weren't specifically designed to run on Apple hardware I'm not sure how you intend to solve that conversion problem without some driver overhead or a translation layer?

It's like complaining that you have to speak English to an English speaker, when you really want to speak German. Apple GPUs speak metal. If you want to speak another language, you're going to need a translator.

Add to that, for things that translate well the translation overhead should be minimal. It’s obviously Vulkan-isms that don’t get hardware acceleration in a metal GPU where you will get the biggest crippling effect. (And of course not being to take advantage of parts of the GPU design Metal is itself designed for.)

 

[PlainBelliedSneetch]

This has been my theory all along, that M doesn't stand for Mac but for mobile. So I bet the iMac 30" and Mac Pro chips will be called D1 or something else.

I had the same thought but the one issue is that Mac mini has the M1.

 

[PlainBelliedSneetch]

I have a 2018 i7 that's meeting my needs, but these two reasons are why I'm upending my usual habit of upgrading nor more than once every 4 years. My Intel Mac is portable enough to bring anywhere, but I always need to be near a power source or I can't work on it for long. Its advertised 11 hour battery life is more like 2-4 for me depending on what I'm doing.

So, you want to be unleashed.