San Jose, CA-based startup SiMa.ai today is making three significant product announcements to accelerate its scaling of physical AI. The company’s next-generation Physical AI silicon, called Modalix, is in production and available for customers along with its SoM (system-on-module) and devkit to accelerate production. It is also debuting LLiMa, a software framework built to deploy high-performance LLMs (large language models) and GenAI models on Modalix for physical AI applications.

Modalix is a second-generation MLSoC (machine learning system on a chip) built to lead the physical AI industry by delivering industry-leading performance and accuracy without sacrificing power, supporting LLMs, transformers, CNNs, and GenAI workloads under 10 W. Its flexible, Arm-based architecture with a native GenAI stack enables real-time perception, decision-making, and natural language interaction. It supports essential interfaces such as camera, Ethernet, and PCIe for scaling physical AI across robotics, automotive, industrial automation, aerospace and defense, smart vision and retail, and medical applications.

“With Modalix now in production, we’re accelerating its global adoption,” said Krishna Rangasayee, Founder and CEO of SiMa.ai. “Demand for our Modalix SoM is strong, and we’re excited to launch it worldwide. We’re also introducing LLiMa to simplify LLM deployment on Modalix, making it easier than ever to bring GenAI to Physical AI systems.”

SiMa.ai was founded in 2018 by Rangasayee, who previously worked at Xilinx, Cypress, Altera, and Groq, to pioneer ultra-efficient machine learning system-on-chip platforms that deliver breakthrough performance per watt for edge AI applications. The company’s sparse neural network processing architecture and comprehensive software stack enable customers to deploy advanced AI capabilities across automotive, industrial, and consumer applications while meeting strict power, cost, and latency requirements.

Modalix meets the stringent power, thermal, and reliability demands of embedded deployments thanks to TSMC’s advanced N6 process.

“TSMC is proud to deepen our collaboration with industry innovators like SiMa.ai to deliver advanced SoCs enabled by TSMC’s leading-edge process technology, meeting the rapidly growing demand for Physical AI,” said Sajiv Dalal, President of TSMC North America.

The new Modalix SoM is pin- and form-factor compatible with leading GPU SoMs, enabling drop-in replacement for rapid integration. It is compact and power-efficient, with integrated MIPI, memory, and essential I/O needed to scale physical AI. The platform empowers developers to go from prototype to production quickly with built-in support for leading ML frameworks, SiMa.ai’s Palette software, and a hardware package including integrated cooling, MIPI camera compatibility, and SSD.

The introduction of LLiMa provides a unified on-device framework for running LLMs, LMMs, and VLMs on Modalix with zero cloud dependency. Featuring open-source or custom LLM import, a curated model zoo, and automated quantization/compilation, it optimizes open-source or SiMa-precompiled models into Modalix-ready binaries. It enables agent-2-agent systems, model context protocol, and retrieval-augmented generation—all fully on-device for physical AI.

For this latest announcement, SiMa.ai leveraged Synopsys’ AI-powered EDA suite, broad IP portfolio, and architecture design and emulation solutions to accelerate development and achieve bug-free A0 silicon, enabling faster, more confident production.

“Achieving a successful first tapeout of MLSoC Modalix illustrates the mission-critical role of Synopsys AI-powered design and IP to achieve complex SoC requirements,” said Ravi Subramanian, Chief Product Management Officer at Synopsys. “Together, Synopsys and SiMa.ai are enabling customers to bring their bleeding-edge AI innovations to market faster and with confidence.”

Why this all matters for the automotive industry was illustrated in July, when SiMa.ai announced the next phase of its strategic collaboration with Synopsys. The deal combines SiMa.ai’s energy-efficient ML processing with Synopsys’ automotive IP and design tools to develop advanced chiplet architectures and reference SoC designs optimized for advanced driver assistance systems (ADAS) and in-vehicle infotainment (IVI).

“Automotive AI presents exciting challenges and massive potential,” said Rangasayee, CEO of SiMa.ai. “With Synopsys’ ecosystem and proven design tools, we can fast-track our ML innovation into automotive platforms.”

As vehicles become increasingly autonomous and connected, automakers face mounting demand for intelligent, efficient AI processing, according to SiMa.ai. The collaboration focuses on overcoming key industry challenges, including ultra-low power consumption for EVs and hybrids, real-time processing for safety-critical ADAS functions, scalable designs for diverse performance needs, cost-effective architectures for broad market adoption, and compliance with automotive-grade functional safety.

“Our partnership with SiMa.ai is driven by strong interest from mutual customers to combine Synopsys’ proven automotive IP portfolio and design engineering expertise with SiMa.ai’s innovative machine learning acceleration technologies to deliver solutions that meet the unique demands of automotive AI applications,” added Subramanian.

The expanded collaboration integrates SiMa.ai’s ML simulators into Synopsys’ design platforms Platform Architect for exploration of architectural options and model ML requirements optimized for automotive OEM specific ADAS/IVI workloads; Virtualizer development kit (VDK), enabling early software development and testing across the automotive ecosystem; and ZeBu emulation, providing pre-silicon power, performance, and efficiency validation, supported by 95-97% accuracy between pre-silicon power emulation estimates and actual silicon.

Machine learning accelerator IP and associated software will be available from SiMa.ai for early access customers by mid-2026, with production release targeted for the end of 2026. A machine-learning IP chiplet, that integrates technologies from both companies, will be available by mid-2027.