The industry’s first multi-modal, software-centric edge AI platform delivers scalable performance per watt and supports a wide range of AI models, from CNNs to multimodal Generative AI, and everything in between.
SiMa.ai has introduced its new product line, MLSoC Modalix, marking the industry’s first multi-modal edge AI product family. The platform supports a range of advanced AI models, including convolutional neural networks (CNNs), transformers, large language models (LLMs), large multi-modal models (LMMs), and generative AI, all optimized for edge applications. The platform offers more than 10 times the performance per watt compared to alternatives, making it a leading solution in terms of power efficiency and performance.
Generative AI’s growing influence is reshaping how humans and machines collaborate, with multi-modal capabilities enabling systems to process various inputs such as text, images, audio, and video. The platform aims to support this transformation across industries like agriculture, logistics, healthcare, defense, and transportation. It enables developers to deploy highly efficient, responsive AI solutions at the edge, while maintaining a compact form factor and minimizing power usage.
The key features include:
- Integrated image signal processor (ISP) module
- Eight lanes of PCIe Gen 5
- Eight Arm Cortex-A65 dual-threaded CPUs
- Built using TSMC’s N6 technology
- Contributes to power efficiency
- Enables compact size
It is the second generation of the company’s commercially successful MLSoC line, offering configurations with 25, 50, 100, and 200 TOPS (tera operations per second) and multiple form factors. The platform provides seamless integration with generative AI applications, facilitating end-to-end pipelines on a single chip. Its compatibility with first-generation MLSoC ensures a smooth transition for existing users, while offering extended support for CNNs, transformer models, and other advanced AI technologies. These advancements enable users to accelerate a wide variety of AI models, including Llama2 7B, and process multiple data types, such as language, vision, and speech, with high precision and low power consumption.
The four models each offering different performance levels (TOPS – Tera Operations Per Second) are:
- 25 TOPS: Ideal for drones, robotics, healthcare triage, and smart cameras (depth/texture).
- 50 TOPS: Supports robotics (AGV/AMR), smart retail, cities, industrial, aerospace, and defense.
- 100 TOPS: Suited for similar tasks as 50 TOPS, with higher performance.
- 200 TOPS: Designed for edge inference servers, enterprise vision systems, and AI gateways.
The platform has already garnered industry validation, with partners like Elementary, Hayden AI, and Advantech expressing their enthusiasm for its potential to power future AI innovations. Samples of the platform will be available to customers in Q4 202.For more information, click here.
