SiMa.ai Launches Modalix PCIe Card to Power Edge AI and LLMs
The card enables full AI workloads to run on its onboard MLSoC architecture, reducing reliance on host CPUs and improving throughput.
SiMa.ai has introduced a new PCIe-based hardware solution aimed at accelerating artificial intelligence workloads at the edge, as demand grows for running large language models (LLMs) outside centralised data centres.
The Modalix PCIe HHHL Card is designed for industrial PCs and edge servers, enabling real-time AI inference in power-constrained environments. Built in partnership with Advantech, the card is part of SiMa.ai’s broader push into “Physical AI” applications across industries such as manufacturing, transportation and defense.
The company said the new card doubles the performance of its predecessor while supporting complex multimodal models and LLMs. Operating at under 10 watts, it is engineered to deliver high efficiency without requiring additional power connectors or active cooling, making it suitable for compact industrial deployments.
“The Modalix PCIe HHHL Card breaks down integration barriers by bringing Physical AI capabilities to both new and legacy IPC infrastructure at exceptionally low power. With direct GMSL camera interfaces and Ethernet connectivity, the entire application runs on the Modalix MLSoC, freeing the host system and delivering the performance, channel density, and power efficiency to meet our customer’s demand,” said Durga Peddireddy, VP Product Management & Partnerships at SiMa.ai.
“In a standard HHHL PCIe form factor, it operates entirely within the PCIe slot power budget — requiring no auxiliary power connector or dedicated active cooling. It also integrates seamlessly into our industrial IPC platforms to deliver efficient, real-time AI processing for real-world applications,” said Brian Wilson, AVP of Sales at Advantech.
The card enables full AI workloads to run on its onboard MLSoC architecture, reducing reliance on host CPUs and improving throughput. It also supports up to 16 video channels and integrates camera and Ethernet connectivity, expanding its use in vision-based AI applications.
Samples of the card are currently available, with general availability expected in the second quarter of 2026.
