Exploring the scalable matrix extension of the Apple M4 processor

https://github.com/tzakharko/m4-sme-exploration

by gok • 2 years ago

184 62 2 years ago

freeqaz• 2 years ago

Any comparison with how much faster this is compared with the previous way of doing things on the CPU?

svnt• 2 years ago

Based on my understanding from the description, it is ~8x faster (250 GFLOPS) for vector ops (vs. SVE mode at 31 GFLOPS which is CPU-ish) and 60-100 times faster (e.g. 2005 GFLOPS) for matrix multiplication for single-precision values.

jandrese2• 2 years ago

That's alright, but not mindblowing. How does it compare to doing the same work on a GPU? Is there a particular set of tasks that GPUs struggle with that would be well suited for this? Or is this more a fig leaf over lousy GPU compute support in Apple land?

TinkersW• 2 years ago

What? The article says this thing does 2005 GFOPs, aka 2 TFLOPS, which is decent, but we have had CPUs that could do more than this for a long time now. My Zen2 12 core does about 3 TFLOPs, and a modern 16 core Zen5 can do 8-10 TFLOPS(I'm unsure what clock speed it can maintain with all cores engaged). And that is generally purpose SIMD not specialized matrix stuff(less generally useful).

Apple CPU's kinda suck at vector ops, but they aren't that bad, this thing is only mildly better. I would guess power savings is a big part of why they use this SVE streaming matrix mode.

ein0p• 2 years ago

I’m not sure why they added this feature. All Apple SoCs have far more energy efficient compute than the CPU. This would only make sense for really tiny models which need extremely quick forward pass. For such models the overhead of a GPU or Neural Engine kernel launch would be quite noticeable. But for those the old NEON was already OK, and if not, there also is a dedicated matrix unit there called AMX. Seems kinda random to me.

adrian_b2• 2 years ago

This replaces AMX, it is its successor.

The older Apple CPUs implemented a custom form of AMX that was not standardized by Arm.

Presumably as a result of cooperation with Apple, the Arm ISA now includes a set of instructions with the same purpose like the original Apple AMX.

The newer Apple CPUs have been updated to use the standard Arm ISA, instead of their older proprietary ISA.

In the Apple CPUs, the former AMX and the current SME provide a much higher throughput than the CPU cores, even if lower than the GPU, and a much lower latency than the GPU, even if higher than the CPU cores.

AMX/SME is implemented as a separate accelerator, distinct from the CPU cores, because this saves power and area in comparison with implementing such instructions in each CPU core. The Apple CPUs do not attempt to compete in high-performance computing applications, so the extra throughput provided by a separate shared matrix operation accelerator is good enough for them.

saagarjha• 2 years ago

This has both actually.

Archit3ch2• 2 years ago

Dedicated FP64 is great for real-time audio processing. Like an included DSP chip.

phkahler• 2 years ago

Isn't FP32 sufficient for audio processing? Even though we have 24bit DACs and ADCs these days I feel like 16bit was really good enough. FP32 with 24bit mantissa should avoid rounding errors at the 16bit level right?

GeekyBear• 2 years ago

Matrix multiplication is very commonly used in science and engineering, not just machine learning.

The neural engine is optimized for machine learning use cases.

This standardized successor to AMX is more general purpose than the neural engine and has much improved matrix multiplication performance vs NEON.

As a bonus, since this is no longer just an experimental implementation of a matrix unit, you get documented access to the new ARM standardized low level instruction set.

brigade• 2 years ago

The neural engine by design cannot handle all possible kernels, and the GPU is significantly slower for integer math, and cannot do fp64. Then for the iPhone SoCs with 4 or 5 core GPUs, the GPU is a bit slower for fp16 and fp32 too.

lxgr5• 2 years ago

if not, there also is a dedicated matrix unit there called AMX

This seems to be the successor to AMX.

dividuum• 2 years ago

Although Apple has included a matrix accelerator in its devices since 2019, it used a proprietary instruction set inaccessible to developers, who officially could only use Apple-provided numerical libraries.

How does that work? Does the hardware throw some kind of fault when using those instructions? Or are they merely undocumented and you could use them if you figure out how they work? I guess the second, as hinted by the "officially"?

my123• 2 years ago

Merely undocumented

jonstewart• 2 years ago

Peter Cawley has a good write-up on the undocumented M1/M2/M3 AMX instructions: https://github.com/corsix/amx

bee_rider4• 2 years ago

As others have said, just undocumented.

IIRC there was a BLIS fork that used AMX instructions. I think it was unofficial though(?). It is hard to do science without properly documented tools.

nxobject2• 2 years ago

If Apple’s going for one SME accelerator per base M4 chiplet, it’ll be interesting to see how to program scalably for Pro/Max/Ultra variants.

wtallis3• 2 years ago

You should be thinking in terms of CPU clusters, not chiplets. The Ultra is the only one with multiple chiplets, but all of their processors have multiple CPU clusters, and so far it's one AMX/SME per cluster.

bee_rider4• 2 years ago

I guess the the CPU/cluster and cluster/chiplet ratios change from generation to generation?

nxobject2• 2 years ago

Ah, thank you! That’s the right word. They’re on the same die, no, so “chiplet” isn’t the appropriate word?

softwaredoug• 2 years ago

I just wish they’d make native tensorflow installation actually work without a million apple silicon specific exceptions :)

TheFuzzball• 2 years ago

They will, just in time for everyone to have switched to pytorch!

kjkjadksj4• 2 years ago

I wish they made computers that ran software like games again. Seems like the last few iterations they’ve been working hard on making computers that are able to run ai models a little faster. Are people really asking for that? I would think far more people would like to play a video game over rolling their own matrix multiplication, but I guess that’s why they pay the people at apple the big bucks because they must know best.

jwells892• 2 years ago

Overall, GPU strength is the best it's ever been in portable Apple devices by a significant margin. The problem isn't the hardware, it's that game developers are reticent to support anything that's not x86 Windows+DirectX or one of the consoles.

It's often said that macOS/iOS supporting Vulkan would help and while I think that's true to an extent, native Vulkan support is still rare enough that it's not going to change all that much in terms of ease of porting. It might improve things on the front of running games through WINE (DirectX → Vulkan translation), but unless developers produce ARM builds of their games there's always going to be the overhead of being run through an x86 translator, which varies depending on how CPU heavy the game is.

wkat42422• 2 years ago

Anything other than metal will be a big win. It's just a non-starter for developers of desktop quality games. Mobile ports sure.

But Apple has never really cared about gaming. During the powerpc era they had a short phase of paying aspyr to make some ports (most notably CoD 4 modern warfare and some battlefield ports) but it was over within a year.

Then about a decade later they had a phase around the 320M chipset where they promoted game releases. And again within a year they dropped the efforts and also let their OpenGL go totally stagnant. This caused for example elite dangerous to drop support.

Now we're stuck with metal. Apple is just too small in gaming for desktop game devs to bother with metal. Not sure if Vulkan will be best but metal surely isn't. And the added complexity of building for arm doesn't help either (arm on windows is non-existent on any hardware aimed at gaming)

I don't think Apple and Mac gaming will ever really become serious. I'm sure that if they do partner with studios like the last few times they'll just abandon the efforts like they always have.

I think the biggest problem is just Apple's total lack of interest (save for the few half-hearted efforts above) to make Mac gaming real.

samatman• 2 years ago

Are you under the impression that fast matrix operations in the CPU are useless for,, games?

Where did you get that idea?

kjkjadksj4• 2 years ago

What matters for games is apple rebuilding bridges with game devs and creating a developer environment that supports game dev on the platform once again.

naming_the_user• 2 years ago

Modern Apple Silicon based laptops have fantastic graphics performance, manufacturers just aren't that interested in supporting them.

It's probably a bit of a chicken and egg thing at this point, plus the fact that most "serious" gamers are going to have desktop PC's anyway.

Detrytus• 2 years ago

I thought one of the reasons to bring Apple Silicon to Mac was that all the iPhone games can now be easily ported?

lxgr5• 2 years ago

manufacturers just aren't that interested in supporting them

AAA games are starting to show up on Steam for macOS these days. Baldur's Gate 3 runs pretty well, for example!

The real shame is that some older indie games are disappearing just as easily, given Apple's deprecation strategy – while Microsoft basically never breaks backwards compatibility, Apple recently cut off 32 bit games (killing about half my Steam library), and presumably Intel-only binaries are next.

aseipp• 2 years ago

You can spend a small amount of die space on something that will yield 10x performance benefits for some things, and you can spend a lot of die space on something that will only yield a general 5% improvement. Which you choose depends on a lot of factors. In other words, the relationship between the "things on the chip" and general performance, or specific application performance, is not a strictly linear relationship.

The 20 series Nvidia GPUs with RTX were a good example. RT cores were added and took up significant die space, people said "why not more CUDA cores", but given the design of consumer GPUs it's extremely unlikely that just replacing those with more CUDA cores would have had a proportional uplift. In Nvidia's case, they realized RT cores were a better bet and served their customer bases (industrial graphics, gaming) better than just more raw numbers.

As it stands, specialization like this is a key element of new designs on leading edge processes. You're going to see more of it, not less.

I guess that’s why they pay the people at apple the big bucks because they must know best.

Well I don't know about "best", they almost certainly know ~infinitely more about their customers and workloads than random people like us do, I can at least say that much.

wmf• 2 years ago

The 20 series Nvidia GPUs with RTX were a good example. RT cores were added and took up significant die space, people said "why not more CUDA cores"

Or they could have had the same number of CUDA cores without RT at a lower price (the fabled "1180")...

fragmede• 2 years ago

They are! The graphics for video games are just a series of matrix multiplications. Before it can get shown to the screen, the graphics are a bunch of triangles, represented by matrices, and in order to do anything in game, those matrices need to be multiplied in order to move them around in 3d space, before getting rendered out to the screen. Making computers better at matrix math means better rendering for video games.

kjkjadksj4• 2 years ago

Apples issue is not whether it is a powerful enough device to run games, but an issue of software compatibility with modern games.

TOMDM• 2 years ago

Apple literally marketed the new iPhone running Death Stranding

lxgr5• 2 years ago

Are you implying that recent Apple SoCs can't run games?

While there's the ML-centric "Neural Engine", the GPU really isn't stagnating by any means: Just in the iPhone 16 presentation this week, ray tracing and a 20% faster GPU were among the headline features. Gaming got its own section in the video presentation!

The fastest GPU I own is in my Mac; the second fastest is in my iPhone. My dedicated (last-gen) game consoles are a distant third and forth, respectively.

kjkjadksj4• 2 years ago

The issue isn’t the hardware its the software.

astrange3• 2 years ago

Higher AI performance actually means higher game performance, because you can render the game at lower resolution and use ML upscaling. Very popular technique now, especially because people prefer higher frame rates over higher resolutions.

brcmthrowaway• 2 years ago

I'm dim, whats the difference between SVE and SME?

mrmuagi• 2 years ago

Vector vs Matrices. Higher dimensional.

DanielLee5• 2 years ago

Great review.

rerdavies2• 2 years ago

In my experience, based on profiling and optimizing of ML-based guitar amp models in the PiPedal project (https://rerdavies.github.io/pipedal/), when using only neon instructions, performance is almost completely constrained by L2 memory bandwidth. Compute cost almost completely disappear while waiting for memory loads and stores.

So, although these devices have ferociously impressive FLOP rates, I'm extremely curious as to how the cost of memory loads and stores is going to work.

I can very well imagine that having large local tile buffers is going to dramatically improve performance. But I'm curious how much. No matter how fast the compute speed is, it seems to me that performance of these sorts of devices in practice is going to be constrained by memory transfer rates. And perhaps by L1 caches in the tile compute unit that are better optimized for tile computation than the L1 cache on a general-purpose cPU.

My current expectation: that performance of matrix multiplies increases linearly with respect to tile size. i.e. a tile size if 8x8 floats will perform twice as fast as a matrix multiplier with a tile size of 4x4, since doubling the tile size reduces the required transfers to and from L2 by a factor of two.

So, compared to a basic A72 ARM neon (effectively, 4x8 tile size), I would expect about a 4x improvement by virtue of the fact that the tile size is larger on the Apple tile processor. Both entirely otherwise limited by the cost of L2 memory loads and stores. And maybe another 2x or 3x improvement because the tile processor L1 caches (tile buffers) are tuned for tile multiply/accumulate operations.

Could somebody comment on how these devices actually perform on real matrix multiplies? It seems inconceivable to me that these devices will actually achieve peak FLOP rates in anything but meaningless test cases. And also somewhat of a meaningless exercise to measure peak performance using test cases that are designed to completely eliminate L2 memory transfers.

Exploring the scalable matrix extension of the Apple M4 processor

Related Stories