Tencent recently released Xingmai Network 2.0, an upgrade that significantly improves the training performance of large-scale artificial intelligence models. It has made breakthrough progress in network scale, hardware performance, communication protocols and fault diagnosis, laying a solid foundation for larger-scale AI model training in the future. This upgrade not only supports 100,000 card networking in a single cluster, but also doubles the switch capacity and silicon optical module speed in terms of hardware. It is also equipped with self-developed computing power network cards, and the communication bandwidth reaches the industry-leading level. What is more noteworthy is that the application of the new TiTa2.0 protocol and TCCL2.0 collective communication library has increased communication efficiency by 60% and large model training efficiency by 20%.
It is understood that in terms of network scale, Xingmai Network 2.0 supports a single cluster of 100,000 cards, providing strong infrastructure support for large-scale AI training. This expansion lays the foundation for larger-scale AI model training in the future.
In terms of hardware upgrades, the capacity of Tencent's self-developed switches has been increased from 25.6T to 51.2T, doubling the capacity. At the same time, the rate of self-developed silicon optical modules has been upgraded from 200G to 400G, and the rate has also doubled. The new version is also equipped with a self-developed computing power network card, which brings the communication bandwidth of the entire machine to 3.2T, ranking first in the industry. These hardware upgrades provide a solid foundation for significant improvements in network performance.
In terms of communication protocols, Tencent has launched a new TiTa2.0 protocol, and its deployment location has been moved from switches to network cards. At the same time, the congestion algorithm has also been upgraded to an active congestion control algorithm. These optimizations have increased communication efficiency by 30% and large model training efficiency by 10%.
In addition, Tencent also launched a new high-performance collective communication library TCCL2.0. This library uses NVLINK+NET heterogeneous parallel communication technology to realize parallel transmission of data. Coupled with the Auto-Tune Network Expert adaptive algorithm, the system can automatically adjust various parameters based on differences in model, network size, model algorithm, etc. This upgrade improves communication performance by another 30% and increases large model training efficiency by an additional 10%.
It is worth noting that the superposition of the upgrade effects of TiTa and TCCL has increased the communication efficiency of the Xingmai network by a total of 60%, and the overall large model training efficiency has increased by 20%. This significant performance improvement will greatly accelerate the training process of AI models and provide researchers and developers with a more efficient working environment.
The upgrade of Xingmai Network 2.0 demonstrates Tencent’s leading position in the field of network technology in many aspects. Its significant performance improvement will have a positive impact on the development of the field of artificial intelligence and promote the training and development of larger-scale and more complex AI models. application. This marks an important step for Tencent in building high-performance AI infrastructure.