Open FPGA Toolchains and Machine Learning with Brian Faith of QuickLogic

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
    • Intro to S3
    • Intro to QuickFeather, EOS S3
    • How to Program QuickFeather using TinyFPGA
    • Hello World on 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