→ Back to Home
Edge AI

DeepX Tapes Out DX-M2: A Sub-$50 Edge AI Chip Designed to Run Generative AI Models Locally at 5W

South Korean fabless startup DeepX has announced the tape-out of its second-generation edge AI chip, the DX-M2. Built on Samsung Electronics' 2nm process, this single-die part is designed to run generative AI models locally at approximately 5 watts, with a target price of under $50 per chip. Mass production is slated for the fall of 2027, with initial silicon expected in March 2027 and customer evaluation samples in the second quarter. This announcement is highly significant for practitioners in the embedded systems and AI hardware space. The sub-$50 price point, if achieved, represents a substantial reduction in the bill of materials for devices requiring on-device generative AI capabilities. Historically, integrating powerful AI inference at the edge has been constrained by the high cost and power demands of suitable accelerators. The DX-M2 aims to address this by offering an economical and energy-efficient solution, potentially enabling a new wave of intelligent, battery-powered products. This could empower developers to move beyond simple inference tasks to more complex generative AI applications directly on the device, reducing reliance on cloud infrastructure for real-time decision-making and enhancing data privacy. This move by DeepX aligns with the broader, well-established trend in cloud, DevOps, and AI towards decentralization and edge computing. The industry has been consistently pushing AI capabilities closer to the data source to address issues of latency, bandwidth, privacy, and cost. Companies like Google with its on-device AI expansion and Microsoft with its sovereign edge solutions have been championing this shift. The increasing prevalence of NPUs (Neural Processing Units) in modern devices, from smartphones to industrial PCs, underscores the growing demand for specialized hardware capable of efficient on-device AI processing. The DX-M2's focus on memory bandwidth (a threefold jump from its predecessor) and custom processing-in-memory technology for improved token generation on models like Llama 3.1 8B highlights the critical role of memory optimization in enabling larger models on resource-constrained edge devices. In practice, this means that developers and product managers should closely monitor the performance benchmarks and customer evaluations of the DX-M2 as they become available. If the promised performance and cost targets are met, it could necessitate a re-evaluation of current edge AI strategies and product roadmaps. The ability to run 20-100 billion parameter models locally at low power and cost could unlock new use cases in robotics, drones, AI PCs, and industrial IoT, where real-time, offline, and private AI processing is paramount. Practitioners should consider how this could impact their cloud spending, data governance, and the user experience of their intelligent products. The trade-off, as always, will be balancing the desire for cutting-edge local AI with the practicalities of deployment and ongoing model updates in a distributed environment.
#edge ai#ai chips#generative ai#low power#hardware acceleration
Read original source