DSP on Beaglebone AI

Started by Crandles 2 weeks ago3 replieslatest reply 2 weeks ago111 views

Has anyone had success or experience in this? I've seen a lot of stuff for the BeagleBone Black (including the extension) but I'm not sure if anything has progressed for the latest iteration.

[ - ]
Reply by jbrowerSeptember 9, 2020


Unfortunately the general impression is that TI has bailed on AI. Maybe that's not fully accurate but it is accurate to say for today's data scientists and ML developers TI is a bygone era.

It would have been good if TI had fully embraced servers, PCIe cards with their multicore DSPs, mainstream deep learning based open sources like Kaldi, and otherwise kept up with the times. Continuing to push small EVM boards, JTAG emulators, and Windows IDEs does not attract third party and ecosystem energy.


[ - ]
Reply by CrandlesSeptember 10, 2020

Hi Jeff,

Are you saying this wouldn't be a good platform to experiment with DSP algorithms and hardware on? I'm not too concerned with the AI label, I was mainly interested in it as it has two C66x DSPs on it.



[ - ]
Reply by jbrowerSeptember 10, 2020


As with all chips and evm boards TI makes, it's an excellent piece of hardware -- amazingly small, high performance, low power.  Assuming TI's code development tools are solid for it (no reason to think not based on prior experience), then it would be an great platform for experimenting with DSP algorithms, but so would any generic x86 server.  If you have a specific application in mind that combines DSP with USB, GbE, video, or audio interfaces then indeed it might be the only suitable choice available.

My general problem with TI's stubborn evm-centric approach is you can't plug those into servers and integrate into mainstream open source code bases for HPC, speech / image recognition, NLP, etc. EE students don't study DSP anymore, they study data science and machine learning.

For practicing engineers, if you run into issues with the Beaglebone you're stuck on e2e island.