Noman Hashim, CEO of Lemur Imaging, demonstrates the company’s latest edge AI and vision technologies and products at the December 2023 Edge AI and Vision Alliance Forum. Specifically, Hashim demonstrates Lemur Imaging’s high-quality 4-bit line memory reduction (LMR) image compression technology, which outperforms 4-bit quantization in edge AI subsystems.
The image quality delivered by Lemur’s innovative compression technology preserves the necessary data to ensure that popular vision detection models, including YOLO and ResNET, give the same level of detection quality using LMR4 data as with INT8, as well as much higher detection quality as compared to INT4. The advantages of moving to 4-bit data at the edge also result in lower power consumption and lower required silicon area, since both bandwidth and on-chip memory are reduced by 50%.
Lemur Imaging has developed the most compact, high performance compression core in the industry, with core sizes under 10K ASIC gates and latency of just a few clock cycles, thus making it suitable for low cost, high performance edge AI SoCs. The IP core is available now for SoC integration, and customers can evaluate the IP via a bit-exact software simulator that comes preloaded with networks both with and without LMR, allowing them to test the IP with their own data sets.