Computing Power

Decentralized computing power entails computing nodes owned by different entities and organized via the internet. Large models consuming extensive data require increased network bandwidth, posing a significant challenge. Moreover, large models typically utilize multiple AI accelerators for computation. While Nvidia leverages PCIe to achieve 800GBps data transmission across chips, decentralized computing nodes connected via the internet face bandwidth limitations of 100GBps or lower, sometimes unreliable. Bandwidth emerges as a critical issue in the era of large models.