Lattice Ultra-Low Power Solution Brings AI to the Edge of the IoT

June 04, 2018


Lattice Ultra-Low Power Solution Brings AI to the Edge of the IoT

The flexibility of an FPGA opens lots of doors when it comes to machine learning and AI. Throw in small size and low-cost, and you've got a real winner.

Machine learning and artificial intelligence (AI) is being addressed by all corners of our industry. For example, last week’s Five Minutes with… podcast featured an executive from Vicor, a power company. Vicor’s Robert Gendron explained how power has a great effect on the advancement of AI.

From another somewhat unexpected source comes another key AI announcement—Lattice Semiconductor. The company unveiled its sensAI, what the company is calling “a complete technology stack combining modular hardware kits, neural network IP cores, software tools, reference designs and custom design services.” The goal of the sensAI platform is to accelerate the integration of machine learning inferencing into broad market IoT applications, such as mobile, smart home, smart city, smart factory, and smart car. It does this by allowing for the implementation of edge computing closer to the source of data. In addition, it could also address some of the analytics needs at the edge of the IoT.

The sensAI technology has the ability to check most, if not all, of the boxes: a power consumption under 1 W, and even down to 1 mW in certain scenarios, package sizes ranging from 5.5 to 100 mm2, lots of interface flexibility (MIPI CSI-2, LVDS, GigE, etc.), and high-volume pricing that shouldn’t stunt growth, in the $1 to $10 range. As you likely already know, handing these operations in an FPGA offers the flexibility to support evolving algorithms, interfaces, and performance levels.

The sensAI stack includes:

  • Modular hardware platforms: an ECP5-based video-interface platform, including Lattice’s Embedded Vision Development kit, and iCE40 UltraPlus-based Mobile Development Platform.
  • IP cores: a convolutional neural-network accelerator and a binarized neural-network accelerator.
  • Software tools: a neural-network compiler tool for Caffe/TensorFlow, and Lattice’s Radiant and Diamond design software.
  • Reference designs: includes face and key-phrase detection, object counting, face tracking, and speed-sign detection.
  • Design services: Lattice has built up an ecosystem of design service partners for most IoT applications.