韩国自研AI芯片，势头正猛

Core Viewpoint - DeepX showcased its first-generation chips at the Embedded Vision Summit, targeting different markets and revealing insights about its next-generation chips for AI devices and autonomous robots [2][3]. Chip Demonstration - The V1 chip, previously known as L1, is a system-on-chip (SoC) featuring a 5-TOPS neural processing unit (NPU) and a quad-core RISC-V processor, designed for edge devices and priced under $10. It operates at 1-2 watts and can run YOLO v7 at 30 frames per second [3]. - The M1 chip is a larger accelerator with a 25-TOPS NPU and a power consumption of 5 watts, capable of running YOLO v5 for pose estimation in collaborative robot environments [3]. - The H1 multi-chip PCIe card utilizes the M1 chip to run 62 channels of YOLO v7, with the production version expected to use four M1 accelerators due to main processor bottlenecks [3]. Quantization Technology - DeepX's quantization technology aims to migrate algorithms from high-power GPUs to its NPU without sacrificing accuracy. The company focuses on minimizing accuracy loss during the quantization process from FP32 to INT8 [4]. - The quantization algorithm has shown to reduce overfitting, resulting in models that generalize better, sometimes achieving higher accuracy with INT8 than with FP32 [4][5]. Intellectual Property Strategy - DeepX holds 60 patents and has filed 282 more, claiming to have more than any other AI chip company. The company is currently collaborating with around 100 clients across Europe, North America, and Asia [5][6]. Automotive Sector Engagement - DeepX is progressing in licensing its NPU to automotive clients, particularly in Europe and Japan, as OEMs shift strategies to engage directly with chip manufacturers [6]. - There is significant demand for domestic NPU IP in China, as local manufacturers must use domestic chips by 2027 but lack advanced NPU technology [6]. Chip Roadmap - The next-generation silicon chip will support Transformer models, with plans to launch by the end of next year, enabling large language models (LLM) to run on edge devices [7]. - The upcoming V3 chip, redesigned based on customer feedback, will feature a dual-core 15-TOPS NPU and a quad-core Arm Cortex A52 processor, with an average power consumption of under 5 watts [8][9]. Future Developments - DeepX will continue to offer both RISC-V based V1 and Arm based V3 chips, with V3 samples expected to be available by the end of 2024 [9].