Microsoft has applied for trademark protection for deep learning optimization library deepspeed. The trademark, filed on December 8, 2020, mentions “temporary use of online non downloadable computer software for artificial intelligence processing and deep learning”, indicating that Microsoft may be considering offering deepspeed through online cloud services. < p > < p > the library, released in February this year, can train large-scale artificial intelligence models on moderate hardware. It only needs 800 NVIDIA V100 cards to create a deep learning model with trillions of parameters, which is more than five times the largest model in the world. Without deepspeed, the same task requires 4000 NVIDIA A100 blocks (2.5 times faster than V100), and the event is shortened by more than 100 days. < / P > < p > this tool is particularly useful when the gpt-3 display model continues to improve with larger neural networks and datasets. However, it does have the potential to push the latest innovations out of the reach of smaller teams with fewer resources, but deepspeed has allowed these teams to compete with larger services with minimal hardware. Moving to cloud processing may make AI training for large models easier to obtain. After 12 years, “world class Super project” Shantou Bay Tunnel ushers in a historic breakthrough today