I did a deep-dive into RTML on Twitter[0][1] just now after seeing this post, and I hope to bring some context and clarity to this project, which is done in conjunction with the National Science Foundation (NSF).
In brief, Professor Sachin Sapatnekar (ACM Fellow [2016], University of Minnesota, College of Engineering, Department of Electrical and Computer Engineering) has been awarded a $2.2M grant from DARPA, “to build open-source hardware generators for a range of machine learning algorithms that process data in real time.”
> program seeks to create no-human-in-the-loop hardware generators and compilers to enable fully automated creation of ML Application-Specific Integrated Circuits (ASICs)
Of course it's a fairly obvious use of ML/AI. There are just so many fascinating ways this can play out in the future - it's going to be an interesting ride!
Note also, Olofsson is the founder of Adapteva [i].
Interesting, I hadn't heard the Parallela name in years. I remember them making some utterly unrealistic promises about close to triple digit cpu core sbc's back in the first half of the 2010's and that was that.
It’d be nice if they implemented TCP over that 400Kbps w/ 60% accuracy. Would be a nice projection of what sort of loss and efficiency should be expected.
I guess they want silicon for different uses. Embedded with low power consumption (on a soldier/handheld for example), in a vehicle, in a datacenter, etc.
With each situation, your need for precision can vary, and along with it the requirement for processing power.
A model tuned for 75% accuracy is lighter and faster to run (fewer layers) vs 85% (several times bigger, slower). I've made up the numbers to illustrate a point. This is anecdotal and I'm not an ML expert. But here's some evidence:
Only related to ‘low-power’ AI I recently came across Ogma[1] And though the combination of online learning and something that works on a Raspberry Pi was fascinating. Surprised it’s not appeared on here before
17 comments
[ 617 ms ] story [ 1261 ms ] thread@DARPA Twitter thread [2019] on RTML
https://twitter.com/darpa/status/1108760215611236353
Report in OP’s linked PDF authored by Andreas Olofsson
https://d60.darpa.mil/speakers/MrAndreasOlofsson.html
DARPA RTML program overseen by Serge Leef
https://www.darpa.mil/staff/mr-serge-leef
Electronics Resurgence Initiative (ERI) Summit 2019 talk on RTML
https://www.youtube.com/watch?v=zQfddARnNB0
ERI Summit site
https://eri-summit.darpa.mil/
ERI @ DARPA
https://www.darpa.mil/work-with-us/electronics-resurgence-in...
National Science Foundation (NSF) RTML program
https://www.nsf.gov/funding/pgm_summ.jsp?pims_id=505640
My unrolled Twitter thread, for convenience
[0] https://threadreaderapp.com/thread/1292268278958694402.html
Direct link to my Twitter thread
[1] https://twitter.com/aspenmayer/status/1292268278958694402
Edit:
This late edit was informed by additional info about those awarded grants from DARPA to study RTML, from HN user WillSlim95.
https://ece.umn.edu/prof-sachin-sapatnekar-to-lead-darpa-fun...
In brief, Professor Sachin Sapatnekar (ACM Fellow [2016], University of Minnesota, College of Engineering, Department of Electrical and Computer Engineering) has been awarded a $2.2M grant from DARPA, “to build open-source hardware generators for a range of machine learning algorithms that process data in real time.”
His faculty page
https://ece.umn.edu/directory/sapatnekar-sachin/
Quirky OG Web 1.0 faculty page
https://people.ece.umn.edu/users/sachin/
https://ece.umn.edu/prof-sachin-sapatnekar-to-lead-darpa-fun...
Of course it's a fairly obvious use of ML/AI. There are just so many fascinating ways this can play out in the future - it's going to be an interesting ride!
Note also, Olofsson is the founder of Adapteva [i].
[i] https://en.wikipedia.org/wiki/Adapteva
200uW for 400Kbps throughput with 0.6 accuracy
200W for 400Gps with higher accuracy with 0.99 accuracy
I guess will be interesting where in this range the result (if any) will fall.
With each situation, your need for precision can vary, and along with it the requirement for processing power.
A model tuned for 75% accuracy is lighter and faster to run (fewer layers) vs 85% (several times bigger, slower). I've made up the numbers to illustrate a point. This is anecdotal and I'm not an ML expert. But here's some evidence:
1. https://ai.googleblog.com/2019/03/an-all-neural-on-device-sp...
2. https://ai.googleblog.com/2018/05/custom-on-device-ml-models...
3. https://krisp.ai/blog/how-we-shrunk-dnn-to-run-inside-chrome...
[1] https://ogma.ai/category/ogmaneo/
Does anyone get to be a DARPA Program Manager while being a CEO of another company?