Latest Release of Lattice sensAI Solutions Stack Delivers up to 6X Performance Boost on Award-Winning CrossLink-NX FPGAs
Upgraded CNN Engine, NN Compiler, and New Reference Designs Enable 60 Frames-Per-Second Video Streams While Consuming Mere Milliwatts of Power
“By taking advantage of the hardware programmability of FPGAs and tracking the latest innovations in AI and ML technology, we’ve extended the performance capabilities of our sensAI solutions stack to new levels,” said
The latest enhancements to the Lattice sensAI solution stack (v3.1) include:
CNNengine IP – optimizations to the way DSPs are used in the IP block and other architectural features of CrossLink-NX FPGAs deliver heightened CNNperformance to the sensAI stack. For example, when running the stack on a CrossLink-NX FPGA, the Mobilenet CNN model (version 1) can process video data at 60 frames-per-second at a resolution of 224x224, a 6x increase in performance from the prior release. Lattice offers an object detection reference design that can enable similar performance in applications like automated defect detection on an industrial assembly line or presence detection in security cameras.
CNNcompiler – the software compiler used in the sensAI stack is specifically optimized for DSPs implemented on a Lattice FPGA fabric to reduce operations-per-cycle by 50 percent. Additional NN model compression innovations include removal of duplicate weights to reduce memory usage while preserving accuracy.
- More development board options - a new sensAI development board featuring the CrossLink-NX 40K FPGA, an image sensor, microphones, and expansion connectors for interfacing with additional sensors. This board supports the reference designs referred to above and is intended to speed time-to-market for AI/ML applications.
- New end-to-end reference designs and hardware - a low power gesture detection reference design to accelerate customer implementation of contactless human/machine interface (HMI) systems that are gaining popularity in the post-COVID-19 marketplace. The application uses an ultra-low-power Lattice iCE40™ UltraPlus FPGA so developers can implement it in a small form factor using minimal power. The reference design could be a compelling option for developers looking to add HMI to an industrial robot or a battery-powered smart toy.
For More Information
To learn more about the Lattice technologies mentioned above, please visit: