DeepX, a South Korean AI semiconductor company, has announced a major milestone for its DX-M1 chip, achieving a 90% yield rate through Samsung Foundry’s 5nm manufacturing process.
The DX-M1, specifically designed for edge computing and Artificial Intelligence of Things (AIoT) applications, is set to enter mass production by the end of 2024, report Korean sources.
“DX-M1 products will start mass production with the world’s best technology that satisfies all three core values of semiconductors such as price competitiveness, computational performance, power consumption, and heat generation control,” said Kim Nok-won, CEO of DeepX.
The DX-M1 represents a breakthrough in energy-efficient, real-time AI processing, filling a critical gap for industries requiring on-device intelligence. DeepX positions the chip as a tailored solution for decentralized AI, moving away from cloud-dependent architectures.
Real-Time Edge AI for Decentralized Applications
The DX-M1 is engineered for real-time AI processing in energy-constrained environments, making it ideal for physical security, robotics, and smart manufacturing. Its ability to process over 16 video channels simultaneously at 30 frames per second highlights its suitability for intelligent video analytics, machine vision, and industrial automation.
By supporting advanced AI models such as YOLOv9 and vision transformers, the chip caters to a wide range of inference tasks, from object detection to image recognition.
DeepX has emphasized its unique contributions to edge AI, leveraging proprietary technologies like IQ8 compression, which enhances model efficiency without sacrificing accuracy.
Additionally, Smart Memory Access technology significantly reduces memory usage, enabling high performance with Low-Power Double Data Rate (LPDDR) memory while avoiding the costs associated with high-bandwidth memory (HBM).
“We will continue to challenge DeepX products to be on the list of luxury goods by continuously improving usability, portability, and quality of software technology by supplying all application systems that exist in the market,” said Kim Nok-won.
Outperforming NPUs in the Edge AI Space
DeepX’s DX-M1 challenges established AI chips and neural processing unit (NPU) offerings from major players like Nvidia, Apple, and Google.
While Nvidia’s H100 dominates raw performance with up to 4,000 TOPS for INT8 operations, it is designed for data centers and lacks the power efficiency needed for edge deployments. Similarly, Google’s TPU v4 provides 275 TOPS but is also optimized for cloud environments, making it not relevant for edge use cases.
The DX-M1 delivers 25 TOPS while consuming just 5 watts of power, translating to an industry-leading 10 TOPS per watt—far surpassing Tesla’s NPUs, which operate at 2 TOPS per watt.
Apple’s Neural Engine in the M3 chip offers 38 TOPS, but as part of a system-on-chip (SoC), it serves a different purpose than the dedicated architecture of the DX-M1.
The DX-M1’s power efficiency and compatibility with diverse AI models enable it to address unique edge AI challenges, such as latency-sensitive applications in smart factories, surveillance systems, and intelligent mobility.
Advanced Manufacturing and Flexible Applications
The DX-M1 is built on Samsung Foundry’s 5nm process, one of the most advanced manufacturing nodes available today. This process delivers an optimal balance of energy efficiency and computational power, placing DeepX ahead of chips relying on older nodes, such as Google’s TPU v4 and Amazon’s Inferentia2.
For less demanding applications, DeepX employs Samsung’s 28nm node, offering cost-effective solutions without compromising quality. This flexibility ensures that DeepX can address a broad spectrum of market needs, from industrial robots to consumer-grade AIoT devices.
Growing Market Presence and Industry Recognition
DeepX has introduced the DX-M1 to over 40 global customers through its Early Engagement Customer Program, which enables partners to trial its chips in diverse applications.
The company’s innovations have been recognized with CES Innovation Awards in Embedded Technology and Robotics categories, underscoring its potential to redefine edge AI hardware. With the global market for AI video analytics and AIoT projected to grow to $75.35 billion by 2028, the DX-M1 aims to play an important role.
