I wonder if Apple ever followed up with this: https://github.com/apple/ml-ane-transformers

They claim their ANE-optimized models achieve "up to 10 times faster and 14 times lower peak memory consumption compared to baseline implementations."

AFAIK, neither MLX nor llama.cpp support ANE. Though llama.cpp started exploring this idea [0].

What's weird is that MLX is made by Apple and yet, they can't support ANE given its closed-source API! [1]

[0]: https://github.com/ggml-org/llama.cpp/issues/10453

[1]: https://github.com/ml-explore/mlx/issues/18#issuecomment-184...

Whisper.cpp has a coreml option which gives 3x speed up over cpu only according to the docs: https://github.com/ggml-org/whisper.cpp?tab=readme-ov-file#c...

Some outdated information about bare-metal use of the ANE is available from the Whisper.cpp pull req: https://github.com/ggml-org/whisper.cpp/pull/1021 Even more outdated information at: https://github.com/eiln/ane/tree/33a61249d773f8f50c02ab0b9fe... In short, the early (M1/M2) versions of ANE are unlikely to be useful for modern LLM inference due to their seemingly exclusive focus on statically scheduled FP16 and INT8 MADDs.

More extensive information at https://github.com/tinygrad/tinygrad/tree/master/extra/accel... (from the Tinygrad folks, note that this is also similarly outdated) seems to basically confirm the above.

(The jury is still out for M3/M4 which currently have no Asahi support - thus, no current prospects for driving the ANE bare-metal. Note however that the M3/Pro/Max ANE reported performance numbers are quite close to the M2 version, so there may not be a real improvement there either. M3 Ultra and especially the M4 series may be a different story.)

I wouldn't say that they aren't useful for inference (there are pretty clear performance improvements even from the asahi effort you linked) - it's just that you have to convert the model ahead of time to be compatible with the ANE which is explained in the readme docs for whisper.cpp that I linked above.

I would say though that this likely excludes them from being useful for training purposes.

Note that I was only commenting on modern quantized LLM's that basically avoid formats like FP16 or INT8, preferring lower precision wherever feasible. When in-memory model values must be padded to FP16/INT8 this slashes your effective use of memory bandwidth, which is what determines token generation speed. So the only feasible benefits are really in the prompt pre-processing phase, and even then only in lower power use compared to GPU, not really in higher speed.

That's really interesting! I didn't know that about the padding behavior here. I am interested to know which models this would include? I know Gemma 3 raw is bf16 - are you just talking about the quantized versions of these? Or are models being released purely as quantized versions these days? I know Google just released a QAT (Quantization Aware Training) model of Gemma 3 27b - but that base model was already released.

Models may be released as unquantized (and even then they are gradually shifting towards lower precisions over time), but most people are going to be running them in a quantized version simply because that gives you the best bang for your buck (you can fit more interesting models on the same hardware). Of course this is strictly about local LLM inference, though one may reasonably assume that the big players are also doing something similar.

My understanding is that model throughput is fundamentally limited at some point by the fact that the ANE is less wide than the GPU.

At that point, the ANE loses because you have to split the model into chunks and only one fits at a time.

What do you mean by less wide? The main bottleneck for transformers is memory bandwidth. ANE has a much lower ceiling than CPU/GPU (yes, despite unified memory).

Chunking is actually beneficial as long as all the chunks can fit into the ANE’s cache. It speeds up compilation for large network graphs and cached loads are negligible cost. On M1 the cache limit is 3-4GB, but it is higher on M2+.

I was referring to both the lower memory bandwidth and lower FLOPs. The GPU can just do… more at once? For now. Or is that changing?

I had also assumed that loading a chunk from the cache was not free because I’ve seen cache eviction on my M1, but it’s good to know that it’s no longer as big of a limitation.

also, I’m a big fan of your work! I played around with your ModernBERT CoreML port a bit ago

.. who is running LLMs on CPU instead of GPU or TPU/NPU

Actually that's a really good question, I hadn't considered that the comparison here is just CPU vs using Metal (CPU+GPU).

To answer the question though - I think this would be used for cases where you are building an app that wants to utilize a small AI model while at the same time having the GPU free to do graphics related things, which I'm guessing is why Apple stuck these into their hardware in the first place.

Here is an interesting comparison between the two from a whisper.cpp thread - ignoring startup times - the CPU+ANE seems about on par with CPU+GPU: https://github.com/ggml-org/whisper.cpp/pull/566#issuecommen...

It essentially never makes sense to run on the CPU and you will only ever see enthusiasts doing it.

Yes, hammering the GPU too hard can affect the display server, but no, switching to the CPU is not a good alternative

Not switching to the CPU - switching to the ANE (Neural Cores) - if you read the research papers Apple has released - the example I gave is pretty much how it's being used - small image classification models running on the ANE, alongside a graphics app that needs the GPU to be free.

Oh, yes, I misread! It’s great for that

Not all of us own GPUs worth using. Now, among people using macs... Maybe if you had a hardware failure?

Every apple fanboy with a $10k mac studio.

M3 Ultra has a big GPU with 819 GB/sec bandwidth.

LLM performance is twice as fast as RTX 5090

https://creativestrategies.com/mac-studio-m3-ultra-ai-workst...

> LLM performance is twice as fast as RTX 5090

your tests are wrong. you used MLX for Mac Studio (optimized for Apple Silicon) but you didn't use vLLM for 5090. There's no way a machine with half the bandwidth of 5090 delivers twice as fast tok/s.

Yeah that's probably wrong. But the M3 Ultra is good enough for local inferencing, in any case.

Unless it’s a large model that doesn’t fit in the 5090, bust that’s no longer a $4k macstudio I think.

$4k will get you a 96 GB Mac Studio with M3 Ultra (819 GB/sec).

That's 3x the RAM of the 5090.

> That's 3x the RAM of the 5090

And a bit less than half the bandwidth (saying for completeness).

that's orthogonal to the speed discussion.

also, the GP was mostly testing models that fit in both 5090 and Mac Studio.

Pretty sure they're using the 80 GPU cores available in that case.

They're basically second place behind NVIDIA for model inference performance and often the only game in town for the average person if you're trying to run larger models that wont fit in the 16 or 24gb of memory available in top-shelf NVIDIA offerings.

I wouldn't say Apple isn't serious about AI, they had the forethought to build the shared memory architecture with the insane memory bandwidth needed for these types of tasks, while at the same time designing neural cores specifically for small on-device models needed for future apps.

I'd say Apple is currently ahead of NVIDIA in just sheer memory available - which for doing training and inference on large models, it's kinda crucial, at least right now. NVIDIA seems to be purposefully limiting the memory available in their consumers cards which is pretty short sighted I think.

Not true. It performs 20-30% better than a RTX A6000 (I have both). Except it has more than 10 times the VRAM. For a comparison with newer Nvidia cards, benchmarks say it does substantially better than a 5070ti, a bit better than a 4080, and a bit worse than a 5080. But once again, it got 30 times the vram amount of the mentioned cards, which for AI workloads are just expensive toys due to lack of vram indeed.

Not for inferencing. M3 Ultra runs big LLMs twice as fast as RTX 5090.

https://creativestrategies.com/mac-studio-m3-ultra-ai-workst...

RTX 5090 only has 32GB RAM. M3 Ultra has up to 512 GB with 819 GB/sec bandwidth. It can run models that will not fit on an RTX card.

EDIT: Benchmark may not be properly utilizing the 5090. But the M3 Ultra is way more capable than an entry level RTX card at LLM inferencing.

My little $599 Mac Mini does inference about 15-20% slower than a 5070 in my kids’ gaming rig. They cost about the same, and I got a free computer.

Nvidia makes an incredible product, but apples different market segmentation strategy might make it a real player in the long run.

It can run models that cannot fit on TEN rtx 5090s (yes, it can run DeepSeek V3/R1, quantized at 4 bit, at a honest 18-19 tok/s, and that's a model you cannot fit into 10 5090s..).

Right, that's the $9500 Mac Studio with 512GB RAM and 80-core GPU.

16x the RAM of RTX 5090.

There are two versions of the M3 Ultra

28-core CPU, 60-core GPU

32-core CPU, 80-core GPU

Both have a 32-core Neural Engine.

Can we stop with the derisive “fanboy” nonsense? Most people don’t say “FOSS” fanboy or Linux “fanboy” — but plenty of people here are exactly that. It’s a bit insulting to people that like and appreciate Mac hardware; just because you might not like it doesn’t mean you have to be so dismissive. And that Mac Studio is a very impressive computer — but it’s usually the ones that have never used on that seem to have to most opinions about them.

One update from Apple Research since that one was "Deploying Attention-Based Vision Transformers to Apple Neural Engine" (but not clear the relationship. It doesn't build on ane_transformers, but maybe a sister project for vision?)

blog: https://machinelearning.apple.com/research/vision-transforme...

github: https://github.com/apple/ml-vision-transformers-ane

Not a public follow-up but the iOS 17 speech-to-text model has a clever approach to KV caching that works within the ANE’s constraints (fixed size inputs).

I wrote about it here[0] but the gist is you can have a fixed size cache and slide it in chunks with each inference. Not as efficient as a cache that grows by one each time of course.

[0]: https://stephenpanaro.com/blog/inside-apples-2023-transforme...

Based on the graphs "up to 10 times faster" compares before/after flash attention

This more than anything feels emblematic to me, that Apple executives are brain dead when it comes to software. AI seemingly being a step too far. (in the s/w direction.) While they could at some level grok classical s/w, NN-s are that Terra Incognita where Apple executives can not possibly follow in. It's just too strange and mysterious world for them to be able to effectively decide or execute on anything. Worked (20yrs ago) in an industrial R&D lab of a mostly h/w manufacturer for 3 years. It looked to me that the worlds of h/w and s/w, the mindsets, diverged pretty quickly on all important considerations of "what should happen next".

Huh? Apple ships heaps of deep learning-based products/features, and they’ve done so for years. But I’ll agree with you, they’re currently behind in generative AI.

This sorta reminds me of the lie that was pushed when the Snapdragon X laptops were being released last year. Qualcomm implied the NPU would be used for LLMs — and I bought into the BS without looking into it. I still use a Snapdragon laptop as my daily driver (it's fine) but for running models locally, it's still a joke. Despite Qualcomm's claims about running 13B parameter models, software like LM Studio only runs on CPU with NPU support merely "planned for future updates." XDA The NPU isn't even faster than the CPU for LLMs — it's just more power-efficient for small models, not the big ones people actually want to run. Their GPUs aren't much better for this purpose either. The only hope for LLMs is the Vulkan support on the Snapdragon X — which still is half-baked.

AFAIK Windows 11 does use the NPU to run Phi Silica language models and this is available to any app through some API. The models are quite small as you said though.

AnythingLLM uses NPU

Could you provide a pointer to docs for this? It wasn't obvious from an initial read of their docs.

could find some about it on the 1.7.2 changelog here https://docs.anythingllm.com/changelog/v1.7.2

If you don’t mind me asking, what OS do you use on it?

I use Windows 11. Podman/WLS2 works way better than I thought it would. And when Gitbash was finally ported (officially) - that filled other gaps I was missing in my workflow. Windows ARM Python still is lacking all sorts of stuff, but overall I'm pretty productive on it.

I pre/ordered the Snapdragon X Dev kit from Qualcomm - but they ended up delivering a few units -- to only cancel the whole program. The whole thing turned out to be a hot-mess express saga. THAT computer was going to be my Debian rig.

I always felt that the neural engine was wasted silicon, they could add more gpu cores in that die space and redirect the neural processing api to the gpu as needed. But I'm no expert, so if anyone here has a different opinion I'd love to learn from it.

Eyeballing 3rd party annotated die shots [1], it’s about the size of two GPU cores, but achieves 15.8 tflops. Which is more than the reported 14.7 tflops of the 32-core GPU in the binned M4 Max.

[1] https://vengineer.hatenablog.com/entry/2024/10/13/080000

Not really. That's 15.8 fp16 ops compared to 14.7 fp32 ops (that are actually useful outside AI). It would be interesting to see if you can configure the ANE to recover fp32 precision at lower throughput [1].

[1] https://arxiv.org/abs/2203.03341

Apple GPUs run fp16 at the same rate as fp32 except on phones, so it is comparable for ML. No one runs inference from fp32 weights.

But the point was about area efficiency

I'm not a ML guy, but when I needed to train a NN I thought that the my Mac's ANE would help. But actually, despite it being way easier to setup tensorflow + metal + M1 on Mac than to setup tensorflow + cuda + nvidia on Linux, the neural engine cores are not used. Not even for classification, which are their main purpose. I wouldn't say they are wasted silicon, but they are way less useful than what we expect

Does Apple care about third party use of the ANE? There are many iOS/iPadOS features that use it.

Not really. Apple software uses the neural engine all over the place, but rarely do others. Maybe this will change [1]

There was a guy using it for live video transformations and it almost caused the phones to “melt”. [2]

[1] https://machinelearning.apple.com/research/neural-engine-tra...

[2] https://x.com/mattmireles/status/1916874296460456089

I guess it's a hard choice as it's 5x more energy efficient than GPU because it uses systolic array.

For laptops, 2x GPU cores would make more sense, for phones/tablets, energy efficiency is everything.

You're completely right, if you already have a GPU in a system adding tensor cores to it gives you better performance per area.

GPU + dedicated AI HW is virtually always the wrong approach compared to GPU+ tensor cores

I was trying to figure the same thing out a couple months ago, and didn't find much information.

It looked like even ANEMLL provides limited low level access to specifically direct processing toward the Apple Neural Engine, because Core ML still acts as the orchestrator. Instead, flags during conversion of a PyTorch or TensorFlow model can specify ANE-optimized operations, quantization, and parameters hinting at compute targets or optimization strategies. For example `MLModelConfiguration.computeUnits = .cpuAndNeuralEngine` during conversion would disfavor the GPU cores.

Anyway, I didn't actually experiment with this, but at the time I thought maybe there could be a strategy of creating a speculative execution framework, with a small ANE-compatible model to act as the draft model paired with a larger target model running on GPU cores. The idea being that the ANE's low latency and high efficiency could accelerate results.

However, I would be interested to hear the perspective of people who actually know something about the subject.

At least one link/benchmark I saw said the ANE can be 7x faster than GPU (Metal / MPS),

https://discuss.pytorch.org/t/apple-neural-engine-ane-instea...

It seems intuitive that if they design hardware very specifically for these applications (beyond just fast matmuls on a GPU), they could squeeze out more performance.

Performance doesn't matter. Nothing is ever about performance.

It's about performance/power ratios.

If you did that, you'd stumble into the Apple GPU's lack of tensor acceleration hardware. For an Nvidia-like experience you'd have to re-architecture the GPU to subsume the NPU's role, and if that was easy then everyone would have done it by now.

M1/M2 shared a GPU design, same with M3/M4. So maybe M5 will have a new design that includes tensor cores in the GPU.

Yet Siri is still dumber than a doorknob…

The README lacks the most important thing: how many more tokens/sec at the same quantization, compared to llama.cpp / MLX? It is worth to switch default platforms only if there is a major improvement.

I ran R1-8B for both anemll[0] and mlx[1][2] models on an M4 Max.

Prompt: "Tell me a long story about the origins of 42 being the answer."

anemll: 9.3 tok/sec, ~500MB of memory used.

mlx 8bit: 31.33 tok/sec, ~8.5GB of memory used.

mlx bf16: 27.17 tok/sec, ~15.7GB of memory used.

Memory results are from activity monitor across any potentially involved processes, but I feel like I might missing something here...

[0] https://huggingface.co/anemll/anemll-DeepSeekR1-8B-ctx1024_0...

[1] https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Lla...

[2] https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Lla...

Yeah, I looked all over for a comparison and couldn't find anything in the repo, on their social media, etc. I saw some other comments here that said it's supposed to be "15.8 fp16 ops compared to 14.7 fp32 ops" but that isn't really enough to go on. Maybe when I have the time I'll install their TestFlight app and do some comparisons myself.

Apple is a competitive choice simply because their unified memory allows you to get enough ram that would take multiple Gpus to have enough space to run larger models.

Yes, but their refusal to open up the ANE to third-party models negates that. You can get (or will be able to very soon) a Strix Halo Ryzen AI Max+ 395 able to access 96GB of unified RAM (on a 128GB system) for well under half what you'd pay for an equivalent M4 system from Apple.

Hobbyists don't use ANE on Apple Silicon and they won't use XDNA on Strix Halo either. In both cases the GPU is faster and can access most of system memory.

$2,000 vs. $3,500 isn't well under half either.

Doesn't the M3 Ultra have 3-4x the RAM bandwidth though?

At 3-4x the price.

I'm trying to figure out what the secret sauce for this is. It depends on https://github.com/apple/coremltools - is that the key trick or are there other important techniques going on here?

coremltools is the only way to run on ANE, so less of a trick and more of a requirement.

The tricks are more around optimizing for the hardware capabilities/constraints. For instance:

- conv2d is faster than linear (see Apple's post [0]) so you rewrite the model for that (example from the repo [1])

- inputs/outputs are static shapes, so KV cache requires some creativity (I wrote about that here [2])

- compute is float16 (not bfloat16) so occasionally you have to avoid activation overflows

[0]: https://machinelearning.apple.com/research/neural-engine-tra...

[1]: https://github.com/Anemll/Anemll/blob/4bfa0b08183a437e759798...

[2]: https://stephenpanaro.com/blog/kv-cache-for-neural-engine

But coreml utilizes ANE, right? Is there some bottleneck in coreml that requires lower level access?

Man, Apple's tight grip on ANE is kinda nuts - would love to see the day they let folks get real hands-on. you ever think companies hold stuff back just to keep control, or is there actually some big tech reason for it?

People keep saying this but I'm not seeing the big difference with other NPU varieties. Either way we're still talking about very experimental stuff that also tends to be hardwired towards some pre-determined use case. So I'm not surprised that people are running into problems while trying to make these more broadly useful.

True; everybody's NPU hardware is afflicted by awkward hardware and software constraints that don't come close to keeping pace with the rapidly-shifting interests of ML researchers.

To some degree, that's an unavoidable consequence of how long it takes to design and ship specialized hardware with a supporting software stack. By contrast, ML research is moving way faster because they hardly ever ship anything product-like; it's a good day when the installation instructions for some ML thing only includes three steps that amount to "download more Python packages".

And the lack of cross-vendor standardization for APIs and model formats is also at least partly a consequence of various NPUs evolving from very different starting points and original use cases. For example, Intel's NPUs are derived from Movidius, so they were originally designed for computer vision, and it's not at all a surprise that making them do LLMs might be an uphill battle. AMD's NPU comes from Xilinx IP, so their software mess is entirely expected. Apple and Qualcomm NPUs presumably are still designed primarily to serve smartphone use cases, which didn't include LLMs until after today's chips were designed.

It'll be very interesting to see how this space matures over the next several years, and whether the niche of specialized low-power NPUs survives in PCs or if NVIDIA's approach of only using the GPU wins out. A lot of that depends on whether anybody comes up with a true killer app for local on-device AI.

> It'll be very interesting to see how this space matures over the next several years, and whether the niche of specialized low-power NPUs survives in PCs or if NVIDIA's approach of only using the GPU wins out.

GPU's are gaining their own kinds of specialized blocks such as matrix/tensor compute units, or BVH acceleration for ray-tracing (that may or may not turn out to be useful for other stuff). So I'm not sure that there's any real distinction from that POV - a specialized low-power unit in an iGPU is going to be practically indistinguishable from a NPU, except that it will probably be easier to target from existing GPU API's.

> a specialized low-power unit in an iGPU is going to be practically indistinguishable from a NPU, except that it will probably be easier to target from existing GPU API's.

Possibly, depending on how low the power actually is. We can't really tell from NVIDIA's tensor cores, because waking up an NVIDIA discrete GPU at all has a higher power cost than running an NPU. Intel's iGPUs have matrix units, but I'm not sure if they can match their NPU on power or performance.

Is there a performance benefit for inference speed on M-series MacBooks, or is the primary task here simply to get inference working on other platforms (like iOS)? If there is a performance benefit, it would be great to see tokens/s of this vs. Ollama.

See my other comment for results.

mlx is much faster, but anemll appeared to use only 500MB of memory compared to the 8GB mlx used.

I am curious if anyone knows if the neural cores in apple silicon based machines are at all useful in training? I’ve been using the MLX framework but haven’t seen them mentioned anywhere so I’m just wondering if they are only useful for inference? I know whisper.cpp takes advantage of them in the inference context.

Edit: I changed llama.cpp to whisper.cpp - I didn’t realize that llama.cpp doesn’t have a coreml option like whisper.cpp does.

Well, the TensorFlow port to metal was written by Apple, and it doens't use ANE. If even they have chosen to use only GPU probably the ANE wouldn't help in training. I also heard that the ANE is way less powerful than Apple Silicon's GPU, but I don't have numbers

Maybe a quick side shift - What the heck are apples neural cores good for ? Used for ? Use cases ?

In my understanding, they are intended to be utilized by various apps (and Apple intelligence) for performing different machine learning tasks (inference), in a manner that is unobtrusive to the user and the rest of the system. While a GPU would theoretically be more performant, it could potentially lead to excessive energy consumption, temperature rise, fan noise, and other factors that may not be desirable when performing basic tasks like OCR when viewing an image within an application.

Currently, it is for example used through the "Vision Framework" eg for OCR tasks (for instance, when previewing an image in macOS, it performs OCR in the background using the ANE). Additionally, they are utilized by certain apple intelligence features that are executed locally (eg when I asked writing tools to rewrite this comment, I saw a spike in ANE usage).

They can also be used for diffusion image models (through core ml, diffusers has a nice frontend for that) but my understanding is that they are primarily for "light" ML tasks within an application rather than running larger models (though that's also possible, but they are gonna probably run slower than in gpu).

They are definitely good for local inference as evidenced by the pretty amazing performance increase on Apple silicon when used with whisper.cpp, maybe other frameworks that utilize coreml? I think they’re sorta purpose built for doing matrix math.

They're used pretty extensively by stuff like the Photos app for various ML tasks. Not all AI is LLMs.

Even 1gb model is prohibitively big for phones if you want mass adoption.

The 1B model works on iPhones[0].

See my other comments. anemll appears to use less memory.

[0] https://huggingface.co/anemll/anemll-llama-3.2-1B-iOSv2.0

We'll just have to await next week's oai-agi1-0.4b-a0.1b-iq1_xs.gguf

Getting anything to work on Apple proprietary junk is such a chore.

is everyone just waiting for teh DGX Spark? Are they really going to ban local inference?

What do you mean ban? The bandwidth between macs isn't enough to do inference effectively.

I'm pretty sure you can network Macs together via the latest Thunderbolt standards and get pretty decent performance overall. Sure, it will be a bottleneck to some extent but it's still useful for many purposes.

Yes you can do that and shard a very large model across the devices but it's way too slow so you will get no performance gains beyond being able to run a much larger model at all.