DeepX to Produce DX-M2 NPU with Samsung 2nm Process

South Korea-based semiconductor design company DeepX is set to produce its next-generation edge AI chip, DX-M2, leveraging Samsung's cutting-edge 2nm process technology. Designed for ultra-low power operation, the DX-M2 targets a power consumption of under 5 watts, and a compact chiplet version is also under consideration.

Built on the RISC-V architecture, the DX-M2 is designed to deliver up to 40 TOPS (trillion operations per second) and is optimized for running Transformer-based models like GPT and next-gen AI agents. This enables support for large-scale models with up to 20 billion parameters, making it ideal for deployment in embedded AI applications such as vending machines and humanoid robots with natural language capabilities.

In comparison, DeepX's previous-generation chip, DX-M, was manufactured using a 5nm process, delivering 25 TOPS and optimized for hybrid FP/INT8-based video processing tasks. It is scheduled for mass production in the first half of 2025, with primary applications in AI PCs and edge servers.

According to Tim Park, Director of Strategic Marketing at DeepX, the company utilizes mixed precision to balance computational performance with energy efficiency. "We use hybrid precision models and understand how the compiler assigns 8-bit, 4-bit, and floating-point operations, which allows us to fine-tune performance across workloads," he stated.

To date, DeepX has secured $531 million in funding, supported by Hyundai Robotics. "We are currently scaling up operations," Park added.

DeepX's earlier product, DX-M1, was specifically developed for computer vision applications and has been integrated into collaborative projects with BMW, Mercedes-Benz, and Volkswagen for autonomous driving systems. While the company gained substantial insights into the automotive sector through these partnerships, it is currently shifting focus away from direct engagement in the automotive market. Instead, it is exploring heavy-duty vehicle applications, where the chip aims to offer cost-effective alternatives in segments traditionally dominated by Renesas, Texas Instruments, and NXP Semiconductors.

The company is also offering an SoC variant of the chip, known as V3, which integrates an NPU, image signal processor (ISP), and four ARM Cortex-A53 cores. This version targets static and mid-to-high-end IP cameras, capable of running vision models with 150 to 200 million parameters, and supporting devices priced around $20.

Looking ahead, the DX-M2 marks DeepX's first NPU built on Samsung's 2nm process, with a focus on achieving sub-5W power consumption.