South Korea‑based DeepX unveils DX‑M2 edge AI chip for 100B‑parameter LLMs inference on robots and drones, mass production set for fall 2027.
chips inference robotics
| Source: Techmeme | Original article
South Korea's DeepX unveiled the DX-M2 edge AI chip, designed to run inference for 100B‑parameter language models on robots and drones, with mass production slated for fall 2027.
South Korea‑based DeepX has announced the tape‑out of its second‑generation AI accelerator, the DX‑M2. The chip is designed to run inference for 100‑billion‑parameter large language models directly on edge devices such as robots and drones, and DeepX aims to begin mass production in the fall of 2027.
The move marks a notable step toward bringing the most powerful generative‑AI models out of data‑center servers and into autonomous systems that operate in real‑time, often without reliable connectivity. By handling inference locally, the DX‑M2 could reduce latency, lower bandwidth costs and improve privacy for applications ranging from industrial inspection robots to delivery drones. The announcement also signals South Korea’s growing ambition to compete in the high‑performance edge‑AI market, a space currently dominated by Nvidia’s Jetson line and a handful of emerging ASIC providers.
Industry observers will be watching for detailed performance metrics, power‑efficiency figures and the chip’s software stack. Partnerships with robotics manufacturers or drone operators could accelerate adoption, while any alignment with major AI model providers would validate the claim of supporting 100‑billion‑parameter workloads. The timeline—mass production slated for late 2027—gives DeepX a window to refine the design and secure fab capacity amid a crowded semiconductor landscape.
Next steps include the release of benchmark results, confirmation of manufacturing partners, and announcements of early customers or pilot programs. How the DX‑M2 stacks up against competing edge accelerators will shape its impact on the emerging market for on‑device LLM inference.
Sources
Back to AIPULSEN