DeepX in South Korea is designing an edge AI chip DX-M2 with a power consumption of 5W, which will be manufactured using Samsung's 2nm process. The company is still considering a Chiplet version.
The goal of DX-M2 is to achieve a performance of 4 billion operations per second (TOPS) and will be optimized for running Transformer models such as GPT and AI agents. It will be able to process models with up to 20 billion parameters, such as DeepSeek, thus integrating natural language chat functionality into embedded devices such as vending machines and humanoid robots. The customized RISC-V processor is responsible for scheduling.
Previously, the 25TOPS DX-M1 manufactured based on the 5nm process was also about to enter mass production, and the chip will be put into mass production in the first half of 2025. DX-M2 is optimized for video processing, combining floating-point and INT8 weights, and is currently being applied to AI PCs and edge servers. Tim Park, Director of Strategic Marketing at DeepX, said, "We use mixed precision, so we know how many 8-bit, 4-bit, and floating-point numbers the compiler provides, and thus know which aspects need to be optimized
DeepX has received investment from modern robotics companies so far, with a financing amount of $531 million. We are currently in the stage of scaling up, "said Tim Park.
M1 is designed specifically for visual applications, and we have collaborated with BMW, Mercedes Benz, and Volkswagen to develop autonomous driving vision systems - we will receive AECQ100 Level 1 certification in October and November. We have a good understanding of the automotive market, so at this stage we will not directly target the automotive market. In addition, we will also enter the heavy vehicle market and can collaborate with Renesas, Texas Instruments, and NXP because the prices of other chips are too expensive
The second version V3 of DX-M1 adopts SoC design, which adds four ARM Cortex-A53 cores in addition to the neural processing unit (NPU) and image signal processing module of M1. This chip is designed for static cameras and mid to high end IP cameras, capable of processing visual models with parameters ranging from 150m to 200m. It is priced at approximately $20 and samples will be released in August.
DX-M2 will also release a SoC version. We hope to provide a minimum power NPU with a power consumption of less than 5W, similar to M1. We are the first chip to use 2nm technology and will release samples in August 2026. Now, other companies have started building their models, which will mature next year, so this is a perfect match for us. We are also preparing SoC versions
Humanoid robots are another target area for DX-M2. Humanoid robots require inference on devices, which is another market we are targeting, "said Tim Park. The DX-M2 will support NVIDIA's visual language model, which has been used by robot developers such as Figure.
Figure AI has its own model, they previously used Nvidia and are now looking for their own chip, and DX-M2 may be exactly what they are using
In the long run, DeepX is considering offering DX-M2 in the form of 2nm Chiplets, particularly for the automotive market. Once the Chiplet market matures, we hope to sell DX-M2 chips to automotive original equipment manufacturers (OEMs). We must understand the situation because the market will not be open to everyone, but we want to participate
