Startup Lemur Reduces RAM Requirements

Lemur Imaging’s LMR technology compresses image data by up to 50% with no visually perceptible loss in typical cases. A guaranteed compression ratio lets chips function with smaller image buffers.

Bryon Moyer

Image processing requires extensive on-chip memory for image buffers. The resulting SRAM blocks can occupy a large percentage of the die, so reducing buffer sizes can measurably decrease die area and cost—particularly worrisome factors for security cameras and similar edge systems. The image-compression scheme from Lemur Imaging, which the company has branded LMR, shrinks the SRAM necessary for system-on-a-chip (SoC) image buffers by as much as 60%. Adding just a few clock cycles of latency, the compressors and decompressors consume less space than the SRAM they eliminate.

The lossy LMR scheme applies to line buffers an algorithm with a JPEG-like high-level structure but simpler math. Images compressed by 50% are visually lossless in typical cases; those compressed by as much as 60% can serve in AI inference without decreasing accuracy relative to uncompressed images, although the technique’s effectiveness using convolutional neural networks (CNNs) remains under study.

Lemur was founded in 2022 by CEO Noman Hashim and CTO Slava Chesnokov. Both have a background at Apical, a computer-vision intellectual-property (IP) vendor that Arm acquired in 2016. Lemur is self-funded; the founders have forgone outside capital, as they expect to secure commercial licenses over the next six months. Hashim and Chesnokov are the only full-time employees, but they hire contractors, bringing the team size to roughly five.

Although JPEG can achieve greater compression than LMR, it’s a 2D algorithm that operates on 8x8-pixel blocks. Lemur’s technology uses a fast 1D approach that’s better suited to streaming-image input and the line buffers in an image-processing system. It differs from and won’t replace incumbent compression for DRAM or nonvolatile storage, but it does address a need for working memory.

LMR is available for licensing. Although Lemur has no confirmed customers, it says multiple Tier One prospects are evaluating the technology’s performance and ease of integration.