Sugon Releases scaleFabric Token Acceleration System With IBGDA GPU Networking

In September 2026, Sugon released the scaleFabric Token acceleration technology system. The system is centered on the scaleFabric native lossless RDMA high-speed network and, through coordination among computing, storage and networking, reduces waiting during data transmission and reading to improve large-model operating efficiency. Its IBGDA technology, which enables GPUs to initiate network communication directly, has been validated on a cluster of 100,000-plus cards, achieving the first large-scale deployment of its kind in China.

A token is the basic unit of content processed and output by a large model, and its generation efficiency directly affects model response speed and service capability. Under a distributed architecture, token-related data must be passed frequently between different GPUs, and the speed of data flow affects the computing efficiency of the entire cluster. During training, all compute cards must synchronize gradient data, a synchronous operation in which one slightly slower card forces all other cards to wait, and abnormal latency lowers model FLOPs utilization (MFU). Inference is divided into input preprocessing and token-by-token generation, which produce large amounts of cached data that must be moved between nodes; the first-token wait time and typing speed perceived by users are related to data transmission speed.

"Replacing compute with storage" has become an industry optimization direction, meaning that intermediate cached data already generated is saved and reused to conserve computing power. Storage is therefore added to the data flow chain, and the coordination efficiency of computing, networking and storage affects the upper limit of the entire AI system. In large-scale distributed training, network communication accounts for 30% to 50% of time, and network performance directly affects the overall efficiency of the computing system. With the spread of the MoE architecture, model operation generates large volumes of high-frequency concurrent small-message communication, further amplifying CPU overhead on traditional network paths and making it a performance bottleneck.

Domestic compute cards lag behind top overseas products in process technology and advanced packaging. IBGDA technology allows GPU kernels to control network interface cards directly to initiate communication, bypassing the CPU instruction relay in traditional paths and reducing small-message communication latency. The technology suits the communication scenarios of MoE models with large volumes of concurrent small messages. After multi-scenario testing, combined with full-link optimization, overall system performance can improve by 20% to 30%.

On the storage side, the industry has developed multi-level storage acceleration solutions for different scenarios such as model loading, cache reading and writing, and checkpoint storage, allowing compute units to access remote storage more directly and reducing intermediate data movement. Together these technologies form the technical foundation for storage-compute-network coordination. Sugon adopts an open architecture, and the scaleFabric network has completed adaptation and optimization with products from more domestic vendors, providing a domestic native lossless RDMA network option alongside RoCE and NVIDIA InfiniBand. Relying on scaleFabric, Sugon's self-developed IBGDA technology has been validated on a cluster of 100,000-plus cards.