Apple Neural Engine Internal: From ML Algorithm to HW Registers
blackhat.com
blackhat.com
This may be a dumb question, but would the space not be better used for more GPU cores? Then if an app actually does need to do neural net stuff, just use the GPU anyway? This way you wouldn't have the space the ANE uses idle 99.9% of the time.
Neat little tool
> Does anything actually use the ANE?
I poked around for a few minutes to see if I could get anything. I saw a blip of usage on in Photos when adding a photo, probably for the new text recognition feature. I was able to see a lot of usage in Photo Booth, I guess for face detection for effects. There's definitely third-party apps making use of Core ML, but it's not something I know much about. On the iPhone devices, it's also used for Face ID (as the article mentions), which is probably much more important to Apple.
> This may be a dumb question, but would the space not be better used for more GPU cores? Then if an app actually does need to do neural net stuff, just use the GPU anyway?
The ANE is likely going to be more power-efficient than GPU cores loaded up for similar processing. Could they use the space for more GPU cores? Maybe, but they would have to work it into their memory architecture, support the additional power requirements, and presumably handle other engineering challenges like thermal design. Including the ANE also gives them something else to put in their marketing materials lol.
NVIDIA GPUs for example dedicate a modest fraction of their die to Tensor cores, at the cost of less general-purpose CUDA cores.
This does not give us a lot of confidence.