bholland84
Posts: 10
Joined: Tue Mar 14, 2017 3:43 am

arm ml processor

Wed Mar 20, 2019 10:34 pm

Hello all,

If you haven't heard, someone is building out an ARM processor to do ML.

https://developer.arm.com/products/proc ... -processor

This seems like a really great idea for a HAT. Is anyone willing to build it out or help me build out a HAT for it? I can program but I don't have a lot of experience with hardware.

Thanks!

Heater
Posts: 13922
Joined: Tue Jul 17, 2012 3:02 pm

Re: arm ml processor

Thu Mar 21, 2019 4:16 am

You mean ARM is building an embedded ML processor.

An ML HAT sounds like an neat idea. So far I don't see a link to a "buy" button to get any chips or even a development board. I don't see any part numbers we could order from Digikey or anywhere. With out that there is no way to get started.

I suspect these devices are unobtainium for normal people at this time.

Also, for things like image recognition how is one going to get the image data from the Pi's camera to the ML processor on a HAT in any reasonable time.

At this moment a far more likely way to be getting on with an ML HAT for the Pi is to use Sipeed MAIX-I AI module with ist dual core 64 bit K210 processor and on chip Neural Network and FFT engines: https://www.seeedstudio.com/Sipeed-MAIX ... -3210.html for which chips and dev boards are available very cheaply already: https://www.seeedstudio.com/sipeed

Image recognition at 60fps anyone?
Memory in C++ is a leaky abstraction .

User avatar
Gavinmc42
Posts: 4061
Joined: Wed Aug 28, 2013 3:31 am

Re: arm ml processor

Thu Mar 21, 2019 5:26 am

I suspect these devices are unobtainium for normal people at this time.
So were camera sensors at one time.
With a bit of luck the next Pi will be filled with unobtainium?
But that would be pure speculation which will get this post locked.

You can do some ML stuff already with the Arm Compute Library.
Add that to doe300's OpenCL stuff and you can get close to being useful.
Useful if you are a guru, less so for the average Joe.

Mostly CV/AI stuff seems to sit around 300x300 pixels for 60fps.
I suspect any more resolution is just not needed.
No idea what face unlock is using?

There is lots of hats that could be added to Pi's, even USB external Engines like the Intel one.
Lots of ML code is mainstream PC based so lots of optimisation can still be done.

Give it a few years, ML accelerators will be common in most phone chips and most cpu's.
With Pi's high profile, hopefully the chip vendors are having to be beaten off with sticks and we get RPT's pick of the best?
But that just more speculation ;)

Time will reveal all.
I'm dancing on Rainbows.
Raspberries are not Apples or Oranges

Return to “General discussion”