Artificial intelligence chip startup SiMa Technologies Inc. today announced it’s shipping its second-gen system-on-chip platform specifically designed for multimodal physical AI workloads.

The new chip, named MLSoC Modalix, is intended to be embedded in robots, industrial equipment, vehicles and more.

It’s designed for a system-on-module that matches the pin configuration and form factor of leading graphics processing units, enabling it to serve as the physical “brain” for edge devices running AI models. The chip allows companies to run various AI systems, including large language models such as Meta Platforms Inc.’s Llama models, convolutional neural networks, transformer models, vision language models and other machine learning algorithms.

Physical AI refers to the combination of artificial intelligence and physical systems, including robots, autonomous vehicles and smart devices. This integration enables these systems to perceive their environment, interact with it, and learn from their experiences. It combines AI algorithms, such as LLMs, so that they can process data from cameras and other sensors and then make decisions based on that information.

The technology depends heavily on having low-power hardware that can handle computation at the edge, where small computers can process massive amounts of sensor data and run AI models directly in vehicles or robots without needing to offload them to the cloud. The further away from the action that data needs to be processed, the longer it takes for a round-trip for a reaction from a robot or a machine. This latency is an issue when safety or critical decisions must be made, or when networking is not available.

The company originally debuted the Modalix machine learning SoC last year, but is announcing today that it’s shipping the system and chipset, available immediately.

According to SiMa.ai, the Modalix SoM will help operations address the diverse needs of the industry for different types of AI systems across the edge market by providing a chip that can handle many kinds of models and integrate easily into existing systems.

SiMa.ai said it partnered with Synopsys Inc., a leading company in the field of electronic design and automation, to develop and innovate the new chip, accelerate design and achieve bug-free silicon in record time.

“The development of Physical AI applications requires validated, purpose-built silicon and software that is only possible using the most advanced design solutions,” said Ravi Subramanian, chief product management officer of Synopsys.

In addition to the Modalix SoM, SiMa.ai introduced LLiMa, a unified on-device framework for running LLMs, vision language models and more on board. It enables developers to quickly import both open-source and custom AI models into Modalix-ready binaries, while also supporting numerous libraries for physical AI applications.

Vision language models, or VLMs, combine the capabilities of LLMs with visual reasoning, allowing them to process both text and images, which is fundamental for categorizing components of an image or video and understanding it. Combined with robotics or other systems, VLMs can be used to automate visual inspection of products and parts to identify defects and ensure quality standards, enable robots to understand and execute tasks based on natural language commands, track and monitor products through a supply chain, and more.

“The era of Physical AI is here! With Modalix now in production, we’re accelerating its global adoption,” said Krishna Rangasayee, founder and chief executive of SiMa.ai.

Rangasayee added that demand for the system-on-module has been strong leading up to the worldwide launch. The module and development kit are available from SiMa.ai today for enterprise deployments for 1,000 units, starting at $349 for 8-gigabyte SoM and $599 for 32-gigabyte SoM per unit. The devkit is priced at $1,499.

Image: SiliconANGLE/Microsoft Designer