SiMa AI Unveils ‘MLSoC’, a Software-Centric Machine Learning System-on-Chip Platform for Embedded Edge


Products for the “edge” market include drones, Internet of Things devices and phones. And this market is served by the very large field of artificial intelligence computer chips.

Many companies have received venture capital funding to provide chips for AI in smartphones and other embedded computing uses. AI chip startup showcased its “MLSoC”, a system-on-chip to speed up neural networks while consuming less power. The manufacturer says the new chip, which has already started shipping to customers, is the only component “specifically designed” to handle jobs that rely heavily on computer vision.

The description

The chip consists of many elements that are assembled into a single chip using Taiwan Semiconductor’s 16-nanometer production process. Components include a machine learning accelerator and “Mosaic”, a program devoted to matrix multiplications, the cornerstone of neural network processing.

A stand-alone computer vision processor, video encoder and decoder are among the functional units onboard, along with 4 megabytes of on-chip memory and a plethora of memory access and communication chips, including an interface to the 32-bit LPDDR4 memory circuits. The ARM A65 processor core, which is frequently found in automobiles, is also present.

The chip hardware includes software to facilitate performance tuning and support a wide range of workloads.

Use case

Robots, drones, self-driving cars, industrial automation, and applications in the healthcare and government markets are just a few of the markets targeted by’s solution.

How is it different

Many in-vehicle and mobile competitors face Intellectual property juggernaut ARM, Qualcomm, Intel and Nvidia are competitors in the edge market with AMD, now the parent company of Xilinx. However, these companies have historically focused on larger chips that consume much more power, on the scale of tens of watts, whereas this focuses on smartphones to embedded ones that only consume a few microwatts of power. .

According to its developers, the chip has one of the lowest power budgets of any chip currently available for performing standard tasks such as ResNet-50, the most popular neural network for processing ImageNet tasks from image classification.


Please Don't Forget To Join Our ML Subreddit

Prathvik is an ML/AI research content intern at MarktechPost, he is a 3rd year undergraduate at IIT Kharagpur. He has a keen interest in machine learning and data science. He is excited to learn more about the applications of machine learning in different fields of study.


Comments are closed.