Machine Learning on the Neural Engine (of the M1)?

We are testing if and how we can trigger the Neural Engine on the M1 Macbook Air to do Machine Learning.
00:00 Intro
02:02 Predicting
05:40 Training
07:50 Coremltoools (python)
09:05 Outro

[@ravindranathjha6434]

Mere 100 subscribers!
This channel is way better than those "video-editing" benchmarking channels.
Subbed!

[@devdutt3435]

Finally someone explained how ANE works 😃

[@jrub3314]

Nice job, bro. I appreciate your work!

[@RomanPittroff]

Amazing and best timing, Danke!

[@cgmiguel]

Insightful testing!

[@abderemanesoilihi6061]

Nice videos, keep it up. I hope to hear more performances on the new M1 Pro

[@wealthbyGDP]

Keep going brother! Need more ML videos for CoreML etc.

[@utubekullanicisi]

I think I did see a couple of people succesfully doing neural training on the Neural Engine. I might put those videos under this comment if I come across them again

[@cavenotes8378]

That would be very interesting to see.

[@gighavlex]

U have the videos yet? Or the name??

[@jannik3475]

Tolles Video!

[@il5083]

Hope we'll see more application with the neural engine, maybe eventually tensor flow training.

[@josephlabs]

Thanks. Would you be able to do a comparison to desktop GPU's because im deciding wether or not to build an ML server

[@teeI0ck]

|📹 this video is very well explained, helpful, interesting and insightful. 💡
Muito obrigado for sharing 🤝

[@Ded-Ede]

With ML on the M1 Mac mini, I though the Photos app and the Notes app scanning function would be better at recognizing and identifying faces, places, objects and OCR in images. And faster too. I’m totally naive to all this but that was the impression I got. Why is this not so now or is it something to come?

[@dr.mikeybee]

Very interesting. Apple's architecture and APIs are a real mystery. This helps push back the veil. One thing to mention is that Apple M1 also has two ML Accelerators. I think they're also referred to as an AGX coprocessor.

[@SergioWolf843]

The (ANE) is not individually programmable as the accelerators (AMX) from the A13 Bionic the Chipset has two neural engines with different purposes.

[@AhsanZafar]

is there a way to use ANE in python?

[@_cw_]

Thanks for making this video. Could you also put a link to your evaluation script as well? I'm interested to inspect why doesn't it train on GPU and ANE

[@gighavlex]

U know now?

[@_cw_]

nope@@gighavlex

[@mehregankbi]

I'm thinking how using mac minis as servers for ML deployment could be nice. on the other hand, you could use tensor cores that nvidia makes.

[@gighavlex]

Now (2024) we can train ML with neural engine in mac m1?

[@dr.mikeybee]

BTW, the BNNS API only does inference -- no training.

[@qwertzuiop875]

Is the NPU analog? Maybe the reason why you can't use it for training is because it's not precise enough. Analog chips are usually not precise, but very efficient and fast.

[@swipekonme]

what a revelation

[@dr.mikeybee]

Have you posted this code or a notebook on Github?

[@karuneshpalekar5212]

Would be great if you mention your device specs too

[@user-ew4jl4oo5r]

Thank you! Can you compare tensorflow on m1 to gpu

[@cavenotes8378]

I did that in a previous video: czcams.com/video/eRwt_FXTdmg/video.html

[@winson5159]

How to program the neural engines with C language

[@LeicaM11]

Ich mag Core ML auch, musste aber den Mac mini M1 zurücksenden. Habe jetzt nur das Ipad pro M1 und warte schon so lange auf den Mac mini M1x oder M2.

[@ck8708]

Skeptical on the purpose of the neural engine. If child protection can be done using the neural engines on phones, how can one be sure that it doesn't do something else on a mac?

[@notgabby604]

Maybe in future everyone will use Fast Transform fixed-filterbank neural networks and everyone will have to learn the fast Walsh Hadamard transform.

[@hzlin6655]

it seems it gonna like pain in the a** if you want to build something big....

[@shisoka7187]

I hear some German accent

[@CompilationError]

this isn't training . 😒