XaLogic Launches Its $38 RISC-V Edge AI K210 AI Accelerator HAT for Raspberry Pi Vision, Voice
Featuring CNN, cryptographic, and audio acceleration, plus a Trust-M security chip, the new K210 AI Accelerator is an impressive add-on.
XaLogic has officially opened crowdfunding for its K210 AI Accelerator, a Raspberry Pi HAT sized for the tiny Raspberry Pi Zero which adds acceleration for edge AI vision and voice tasks courtesy of a Kendryte K210 system-on-chip.
First announced late last year as a follow-up to the company's XAPIZ3500 accelerator, the XaLogic K210 AI Accelerator is designed around the Kendryte K210 system-on-chip with a dual-core 64-bit RISC-V with independent floating-point units, a Kendryte Processing Unit (KPU) convolutional neural network accelerator coprocessor, an audio accelerator, cryptographic accelerators, and a dedicated fast Fourier transform (FFT) block.
To this XaLogic has added an Infineon Optiga Trust-M security coprocessor, which includes support for ECC, RSA, AES, and TLS cryptographic operations, and a Lattice Semi iCE40 field-programmable gate array (FPGA) — used as an SPI bridge between the K210 and the host Raspberry Pi, which powers and communicates with the add-on via its 40-pin general-purpose input/output (GPIO) header.
According to XaLogic, the K210 AI Accelerator differentiates itself from rival devices like the Google Coral USB Accelerator and Intel Neural Compute Stick 2 with its on-board security chip, ultra-low 0.3W power consumption, and the release of both the firmware source code and hardware schematics — not to mention its bargain-basement price. At the same time, though, there's a gulf in performance: The 0.5 TOPS of the K210 AI Accelerator is overshadowed by the up-to-4 TOPS available from either of the rival devices — at, admittedly, a considerably higher power draw.
The company has released pre-trained models for its accelerator, which is programmed using a plug-in model, for object detection, face detection, age and gender estimation, voice control, and vibration detection, and promises that additional models will come. For those looking to create their own models, XaLogic advises a separate computer — "preferably with an NVIDIA GPU" — will be required for training, with a conversion tool supporting TensorFlow Lite, Caffe, and "limited support of ONNX format."
The crowdfunding campaign is now live over on Crowd Supply, with the K210 AI Accelerator priced at $38 — though users of Raspberry Pi boards which have the Power over Ethernet (PoE) header in place will need an optional $3 riser to ensure the HAT clears the pins.
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.