China's Sugon said it has launched the scaleFabric Token acceleration system, enabling the first large-scale domestic deployment of IBGDA. Built on native lossless RDMA networking, scaleFabric creates a tightly coupled compute–memory–network token-acceleration channel and supports GPU-initiated network communication via IBGDA; the technology has been validated in clusters at the ~100,000‑card scale. Sugon said scaleFabric has been adapted to DeepEP and other communication frameworks and uses IBG

China's Sugon said it has launched the scaleFabric Token acceleration system, enabling the first large-scale domestic deployment of IBGDA. Built on native lossless RDMA networking, scaleFabric creates a tightly coupled compute–memory–network token-acceleration channel and supports GPU-initiated network communication via IBGDA; the technology has been validated in clusters at the ~100,000‑card scale. Sugon said scaleFabric has been adapted to DeepEP and other communication frameworks and uses IBGDA and storage passthrough to boost token transfer efficiency, offering a new high-speed interconnect option for China’s domestic large AI models.