Yesterday they reverse engineered Apple’s neural engine: https://github.com/maderix/ANE

[migueldeicaza@mastodon.social]

Yesterday they reverse engineered Apple’s neural engine: https://github.com/maderix/ANE

And today people are running the amazing Qwen3.5 accelerated on an m1, this is insane:

[pixel@oldbytes.space]

@Migueldeicaza yeah now I definitely need to upgrade my M1 iMac.

[piyuv@techhub.social]

@Migueldeicaza why wasn’t it open in the first place? Oh right, it’s Apple… pay obscene amount of money for hardware you can’t even fully utilize

[migueldeicaza@mastodon.social]

@piyuv internal APIs are done so you can change the implementation without breaking people - that’s why people do it. Every time you expose things it ties your hands.

[dotsie@mastodon.social]

@Migueldeicaza

It will be interesting if this works on the new 40 and probably 80 core M5 Max / Ultra models.

Are people modifying this to leverage operations for inference? The GitHub is focused on training, curious if this could be leveraged to help speed up some TTFT stuff which would be genuinely useful.

Apple should really expose this in the next OS especially with 40 - 80 of these neural accelerators in their top M5 processors.

[joethephish@mastodon.gamedev.place]

@Migueldeicaza I’m confused, isn’t a model that’s converted to MLX effectively always running on the neural engine? Or it just on the GPU?

[tylercheung@mastodon.social]

@Migueldeicaza I think Apple's AI strategy should really be: don't make their own model but make a software strategy that can beat cuda and leverage apple silicon memory model to beat NVidia (i.e. have OpenAI lining up to buy mac minis)

[surendrajat@fosstodon.org]

@Migueldeicaza wasn't it already reversed by @AsahiLinux people here https://github.com/eiln/ane ? What's the difference exactly?