StarryLink Optical Module

With the surge in AI development, AI training clusters have evolved to a scale of 10,000+ GPUs, resulting in a significant increase in the number of optical modules required. For instance, the 1000-GPU cluster needed for training GPT-3 requires interconnections using 2500 200G or 4000 400G optical modules. To train GPT-4, a 10,000-GPU cluster necessitates interconnections with 25,000 200G or 40,000 400G optical modules. Each time a GPU is added, the number of optical modules increases by an average of 2.5 to 4. Efficient network interconnection is crucial for enhancing the computing power of the system cluster.

However, during the actual training process, it has been observed that the failure rate of optical modules is higher than expected, negatively impacting network reliability. The annual failure rate of optical modules is 4‰, leading to an average interruption in training for a 10,000-GPU cluster once every 3.6 days. The average fault recovery time is 2 hours, resulting in a daily waste of CNY1.4 million in computing power investment.

Huawei's StarryLink optical modules offer customers ultra-reliable, long-distance, and highly secure data center network interconnection experiences. The short-distance optical return loss positioning technology enables precise and efficient identification of contaminated or loose optical modules, achieving minute-level fault locating. Additionally, the optical module channel loss resistance technology enhances the reliability of optical modules by a factor of ten.