Signal

South Korea-based DeepX tapes out DX-M2, an edge AI chip to run inference of 100B-parameter LLMs on robots and drones, targeting fall 2027 for mass production

First reported by Eetimes ·

The signal ●●●● Compiled by AI from Eetimes, Techmeme, Bloomberg and Yonhap News Agency
Why you might care

Edge devices can now run advanced AI models locally without needing cloud connectivity.

What happened

South Korean AI chip startup DeepX has announced the tape-out of its second-generation edge AI accelerator, the DX-M2. This new chip integrates Samsung's processor-in-memory (PIM) technology and is designed to run large language models (LLMs) with up to 100 billion parameters on edge devices like robots and drones. The DX-M2 aims to achieve this with a power consumption of less than 5 W, maintaining the same power budget as its predecessor, the DX-M1, which is currently in mass production for vision applications. DeepX claims the DX-M2 offers a 6.7x performance gain over the M1, with tripled memory bandwidth to 150 GB/s and a total performance of 80 TOPS. The chip also benefits from a doubled NPU power efficiency and manufacturing on a 2nm process node. Mass production is targeted for fall 2027.

What it means

The DX-M2 represents a significant leap in bringing powerful AI capabilities directly to the edge, particularly for complex tasks like running large language models on resource-constrained devices. By integrating processor-in-memory (PIM) technology, DeepX is directly addressing the memory bandwidth bottleneck, a critical challenge for AI inference. This focus on on-device processing not only reduces latency and reliance on cloud infrastructure but also enhances privacy and enables operation in environments with intermittent connectivity.

This development signals a growing trend towards highly specialized edge AI hardware capable of handling sophisticated generative AI models. The ability to run 100B-parameter LLMs on devices drawing under 5W could accelerate the adoption of advanced AI in robotics, drones, and autonomous systems, potentially creating new product categories and use cases. Companies and developers in these sectors should monitor DeepX's progress toward mass production, as it could unlock new possibilities for intelligent, self-sufficient edge devices.

AI-written summary. May contain errors.

AI Robotics