Recently, Zhongke Shuguang released the ScaleFabric Token acceleration technology system to achieve the first large-scale application of IBGDA technology in China. The system is based on a native lossless RDMA high-speed network to build a three-level, tightly coupled “computational memory and transmission” token acceleration superchannel. Among them, IBGDA technology, which supports GPUs to directly initiate network communication, has been implemented on a large-scale scale for the first time in China, and has been verified in a large-scale cluster of 100,000 cards. At present, ScaleFabric has completed adapting communication frameworks such as DeepEP, and has improved token circulation efficiency through IBGDA, direct storage and other technologies, and provided new high-speed interconnection options for large domestic models.

Recently, Zhongke Shuguang released the ScaleFabric Token acceleration technology system to achieve the first large-scale application of IBGDA technology in China. The system is based on a native lossless RDMA high-speed network to build a three-level, tightly coupled “computational memory and transmission” token acceleration superchannel. Among them, IBGDA technology, which supports GPUs to directly initiate network communication, has been implemented on a large-scale scale for the first time in China, and has been verified in a large-scale cluster of 100,000 cards. At present, ScaleFabric has completed adapting communication frameworks such as DeepEP, and has improved token circulation efficiency through IBGDA, direct storage and other technologies, and provided new high-speed interconnection options for large domestic models.