这项互联技术，要超越CPO！

Core Viewpoint - The speed of training large AI models ultimately depends on two concepts: scaling up and scaling out, which involve different physical connection technologies [1][2]. Group 1: Scaling Technologies - Scaling out refers to increasing the number of interconnected AI computers to handle large tasks, while scaling up involves integrating as many GPUs as possible within each computer [1]. - Scaling out relies on photonic chips and optical fibers for data transmission, while scaling up uses copper cables, which are simpler and more economical [1]. - The network density achieved through scaling up is approximately ten times that of scaling out, but the data transfer rates required for high-performance computing are approaching the physical limits of copper cables [1][2]. Group 2: Challenges with Copper Cables - Copper cables face limitations due to the skin effect, which causes high-frequency signals to concentrate near the surface, increasing resistance and requiring thicker cables and more power [7][8]. - Active electrical cables (AECs) have been developed to mitigate these issues, but they increase complexity and power consumption [8]. Group 3: Innovations in Data Transmission - Point2 Technology and AttoTude propose a new solution that combines the low cost and reliability of copper cables with the fine size and long-distance capabilities of optical fibers [2][4]. - Point2 plans to mass-produce cables capable of 1.6 terabits per second using polymer waveguides, while AttoTude is developing similar technology using terahertz frequencies [4][11]. - Both companies claim their technologies can easily surpass copper cables in transmission distance, achieving 10-20 meters without significant loss [4][5]. Group 4: Future Prospects - The wireless technology being developed is expected to be more reliable and easier to manufacture than photonic technology, potentially replacing some copper cables in printed circuit boards [5][14]. - The market for vertical expansion networks is growing, with companies looking to integrate more GPUs while minimizing cooling technology needs [15][16]. - Both Point2 and AttoTude are working on versions of their technology that can be directly integrated into GPUs, which could further enhance performance and efficiency [15][16].