TinyVision.ai's Vision FPGA SOM Brings Compact, Low-Power On-Device Computer Vision to the Masses

With 64Mb of SRAM on-board, the Vision FPGA SOM offers low-power computer vision and on-device ML in a compact package.

Gareth Halfacree
10 months ago β€’ Machine Learning & AI / FPGAs
The compact Vision FPGA SOM draws as little as 10W. (πŸ“·: tinyVision.ai)

Low-power computer vision specialist tinyVision.ai is preparing a crowdfunding campaign for its Vision FPGA system-on-module (SOM), combining an image sensor and Lattice Semiconductor iCE40 field-programmable gate array (FPGA) in a compact board drawing just 10-20mW.

"The Vision FPGA SOM is an iCE40 5k FPGA based System on Module that integrates a low power qVGA vision sensor, three-axis accelerometer/gyroscope and an I2S MEMS microphone in a small form factor (3cm x 2cm)," tinyVision.ai explains of the board's design. "We designed the device with the following thoughts in mind: Overall power consumption should be minimized and in the 10-20mW range to enable battery-powered applications; easy to integrate into a larger system using an interrupt driven SPI host interface, programmable IO voltages to support a glueless HW interface; complexity should be localised to the SOM so that the developer can quickly integrate this device into their system using a breadboard as well as a simple API."

Other key design points include ease of integration both mechanically and electrically, a strong feature set including a healthy chunk of static RAM (SRAM), and support for an open source toolchain β€” removing the dependency on proprietary software common to many FPGA-based development boards. The finished design is capable of image, sound, and motion capture, can trigger on scene changes, motion, and or sound inputs, can log accelerometer data along with images for use in virtual and augmented reality projects, supports audio beamforming, edge-processing, and uses the tinyML machine learning platform for object detection, keyword recognition, and gesture processing on-device.

The board is based around the Lattice iCE40UP5k FPGA, which includes 5k look-up tables (LUTs), 1Mb of RAM, and and a Pixart PAJ6100U6 qVGA monochrome image sensor. A connector is included for an optional Himax HMB010 colour rolling-shutter image sensor or the low-cost OV7670 sensor, while audio is handled by a single Knowles MEMS microphone with option to expand to two for stereo capture and motion data measured by a six-axis Invensense 60289 inertial measurement unit (IMU). There's 4Mb qSPI flash and 64Mb qSPI SRAM on-board, along with an infra-red LED for low-light illumination.

The crowdfunding campaign is set to launch soon for module-only and developer kit bundles, the latter of which breaks out all pins and offers USB connectivity for programming and debug, power measurement via I2C, adds an additional microphone and more LEDs, and includes a PMOD expansion header and small prototyping area.

Pricing has yet to be confirmed, but more information and a link to be kept in the loop is available on the project's Crowd Supply page.

Gareth Halfacree
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.
Related articles
Sponsored articles
Related articles