Open FPGA Toolchains and Machine Learning with Brian Faith of QuickLogic

From The Amp Hour Electronics Podcast The Amp Hour (Chris Gammell and David L Jones)

Welcome Brian Faith, CEO of QuickLogic!

Past guest Tim Ansell introduced us to Brian, from their work together on the open toolchain.
They met at a tradeshow and Brian declined the first time, only to be convinced later.
QuickLogic IP licensing
Brian attended OR Conf in Bordeaux, where they were watching talks and excited by future growth of users of the open toolchain.
Resisted for a year
Brian started at QuickLogic during the "schematic era" (when FPGAs were designed using graphical schematic of logic blocks)
Previously their toolchain
- Worked with early versions of Synplicity, but later switched to using Mentor Graphics Precision
- There was no bundled simulator
- Proprietary Place and Route (P&R)
The new QuickLogic approach is Symbiflow
It's also about the software engineer
A community member ported NuttX to the platform
What did QuickLogic give up, in order to use the Symbiflow toolchain? They had to publish the spec of the bitstream
What could you do with the spec of the bitstream? Why is it secret? Apparently, due to history and a generally closed off ecosystem in FPGAs.
QuickLogic is targetign selling to software engineers, not just FPGA engineers. This has become much easier with python targeting FPGAs (LiteX, Migen)
Software users will help enable more "mass customization"
Making software designs into silicon
Open Hardware Group
RISC V
Global Foundries at Munich
The Artic Pro 2 will be built on the Global Foundries 22FDX, which is their 22 nm process
Hardware/Software partitioning
They're building a test chip
QuickFeather
SensiML is the web-based machine learning toolset. The team came from the Intel Curie group.
SensiML was bought by QuickLogic at the beginning of 2019, but they still offer services for chips outside the QuickLogic portfolio as well.
Chris doesn't think a threshold detect algorithm would be up to the task in many cases.
QuickLogic and SensiML are sponsoring a Hackster contest targeted at projects that will help prevent climate change.
You send in your sensor data, SensiML gives back models you include as a "black box" algorithm
The web interfice allows you to dial in performance algorithms. You can also update the data/model later if you want to tweak based on new data or different parameters.
There is an example data set on github using a PM2.5 sensor
QuickLogic Open Reconfigurable Computing (QORC)
Size of the model depends on perfomance dialed in on the website
The models are set to run on on Cortex-M4, specifically the EOS S3
TensorFlow lite for microcontrollers
APIs for convolution
eFPGA = embedded FPGA
In the case of the EOS S3, it's roughly equivalent to 1000-2000 LUTS
USB in the FPGA without a dedicated (hard) USB core can do USB 1.1 full speed data speeds.
Videos and instructions
- EOS and QuickFeather
- SensiML
  - Endpoint AI without Writing Code
  - SensiML Overview
Building a proof of concept
Community edition of SensiML gives you enough access for entering the contest, trying out models at home (non-commercial).
If you are developing a commercial product, SensiML has commercial subscription prices (Chris thinks they're reasonable, relative to hiring an FPGA/DSP engineer)
Removing the gyro using SensiML
Wrist worn wearables for applications like remote control
Industrial applications
Consumer is still a focus
Gerry Roston talking about data and monitoring for large scale auto manufacturing facilities
They are targeting many of their classic customers in the Automotive / MIL / Aero industry, as well as new ones. They are avoiding the server / datacenter industry.
QuickLogic licenses things like their IP blocks, memory blocks, math blocks for people to design into future silicon.
If you licence IP from QuickLogic (fabric), you will also be able to use Symbiflow for your silicon product.
Interested in learning more and giving it a try? Check out the Hackster contest
EOS S3 page
Great video tutorials