后摩智能发布全新端边大模型AI芯片，CEO吴强：要让AI算力像电一样方便好用

Core Viewpoint - The article discusses the launch of the new AI chip, 后摩漫界®M50, by 后摩智能, which aims to empower the widespread adoption of large models at the edge and endpoint devices, emphasizing high performance, low power consumption, and ease of use [2][5]. Group 1: Product Features - The M50 chip achieves 160 TOPS@INT8 and 100 TFLOPS@bFP16 physical computing power, with a maximum of 48GB memory and a bandwidth of 153.6 GB/s, while maintaining a typical power consumption of only 10W [2][3]. - The M50 chip utilizes a second-generation SRAM-CIM dual-port architecture, allowing simultaneous weight loading and matrix computation, supporting mixed-precision operations [3][4]. - Compared to traditional architectures, the M50 chip offers a 5 to 10 times improvement in energy efficiency, making it suitable for edge devices that require fast computation with low power consumption [4]. Group 2: Product Matrix - The product matrix includes the 力擎™ LQ50 M.2 card, which provides AI capabilities for mobile terminals, supporting inference for 7B/8B models at over 25 tokens/s [4]. - The 力谋®LM5050 and 力谋®LM5070 acceleration cards integrate 2 and 4 M50 chips respectively, providing high-density computing power for single and large model inference, reaching up to 640 TOPS [4]. - The BX50 computing box is designed for edge scenarios, supporting 32-channel video analysis and local large model operations [4]. Group 3: Applications and Benefits - The products can be applied in consumer electronics, smart offices, and smart industrial sectors, enabling local processing without the need for internet connectivity, thus enhancing data security [5]. - In consumer devices, the M50 chip enables local large model inference for smart interactions and content generation, ensuring user privacy [5]. - In smart industrial applications, local computing power allows for real-time analysis and decision-making, keeping production data secure and avoiding cloud transmission risks [5]. Group 4: Future Developments - 后摩智能 is developing the next-generation DRAM-PIM technology, which aims to embed computing units directly into DRAM arrays, potentially achieving over 1TB/s on-chip bandwidth and tripling current energy efficiency levels [5]. - The goal is to make powerful AI computing capabilities widely accessible in everyday devices like PCs and tablets [5][8]. Group 5: Investment and Support - In the past two years, 后摩智能 has received investments from various funds, including the China Mobile Industry Chain Development Fund and the Beijing Artificial Intelligence Fund, supporting ongoing innovation in edge model chip technology [6].