For the new network communication challenges of efficient data interaction between components in open interactive environments, a novel C2N(computing and control network)was proposed.Aiming at the extreme requirements for efficiency, real-time performance, flexibility, and security, C2N adopts intelligent and simplified designs in protocol architecture, planning, application, and security design, providing high-performance and highly flexible basic network support for strong real-time collaborative fusion among heterogeneous resources.Based on a detailed investigation of relevant research work, key technologies of C2N were discussed, such as data link layer enhancement, remote direct memory access for sensor-controllers, and service-oriented sensing and control middleware.It also introduced the key technology research and test evaluation carried out by the network chip and system team of the National University of Defense Technology, and prospected future challenges and research directions to help China gain leading advantages in high-end equipment systems and innovative ecosystems.

To further optimize the hardware offloading of collective communication based on the network interface card in the "Tianhe" network, and to support more types of collective communication algorithms and larger message sizes, the order-preserving triggering mechanism and data buffering method for collective communication hardware offloading was investigated.An order-preserving triggering mechanism for concurrent multitasking was proposed, which meets the desired semantics of collective communication and ensures the reproducibility of floating-point computation results.A dynamic network data buffering method based on Hash tables and pulsed credit flow control was proposed to alleviate the contradiction between limited hardware buffering resources and the high demand for buffering a large amount of network data from concurrent multitasking.Experimental results show that compared with software-based collective communication operations, this method can support the hardware offloading of various algorithms for several typical collective communication operations, with significant performance improvement. Meanwhile, the hardware implementation cost is low, especially with high utilization of buffering resources.

An ER-MKKNN(enhanced random mixed kernel K-nearest neighbors algorithm)was developed to meet the requirements of base station network traffic prediction in ultra-dense 5 G/6G environments.A hybrid kernel function was formed by combining a radial basis function kernel with a white-noise kernel, thereby overcoming the trade-off between nonlinear relationship modeling and noise suppression that plagues single-kernel methods.Dual random subsampling of both samples and features, together with a randomized hyperparameter-interval strategy, was employed to bolster generalization stability in high-dimensional, sparse settings.A dynamic weight-allocation mechanism based on inversion of out-of-bag errors was introduced to improve robustness against abrupt traffic fluctuations.Finally, a multi-level parallel architecture was implemented to deliver a scalable prediction framework for ultra-dense network topologies.Experimental evaluations show that ER-MKKNN outperformed deep-learning models in root mean square error, mean absolute percentage error and mean absolute error, respectively, establishing a new technical pathway for intelligent network operations and maintenance.

Particle transport simulations using stochastic methods face significant challenges on conventional von Neumann architectures, particularly due to random branching events and irregular memory access patterns.These limitations stem from the fundamental mismatch between probabilistic algorithms and deterministic computing paradigms.To bridge the gap between architecture and algorithms, a probabilistically tunable true random number generator was developed based on spintronic and ferroelectric devices.The physical randomness of spintronic devices was leveraged to provide a physical random source for the architecture, and the throughput of random bits was enhanced through optimized control logic and writing mechanisms.Next, programmable synapses were designed based on the memristive properties of ferroelectric devices, enabling nonvolatile continuous weight storage with tunable probabilities.The experimental results indicate that the proposed approach achieves performance improvements ranging from 171 to 1028 times compared to a general-purpose CPU when solving a sample transport problem.Furthermore, compared to existing spin-transfer torque magnetic tunnel junction based true random number generators, the developed method not only enables tunable probability random sampling but also achieves a throughput of 303 Mbit/s when generating uniformly distributed random sequences.

Combining software defined networking and SR(segment routing)can optimize network performance, but in large-scale dynamic networks, excessive link utilization at key nodes can lead to a surge in queue delays.To address this, a SROD-LC(segment routing optimization algorithm based on deep reinforcement learning and load centrality theory)was proposed.By quantifying the importance of network nodes using load centrality theory, key nodes are identified and their link load states are monitored;utilizing a multi-agent reinforcement learning framework, distributed deep reinforcement learning agents are deployed at key nodes, coordinating routing decisions through a shared reward mechanism to achieve proactive optimization of link loads.At the same time, leveraging the flexibility of SR, segment identifier lists are dynamically adjusted to quickly reroute partial traffic, reducing local link utilization and avoiding potential congestion.Simulation experiments based on real network topologies show that when the proportion of SR key nodes is in the range of 0.3~0.5, the SROD-LC algorithm exhibits significant optimization effects, reducing the network′s maximum link utilization by 21%~35% compared to baseline algorithms.

Efficient inference deployment of large language models faces severe challenges in resource-constrained scenarios.Although current mainstream inference optimization techniques have improved model inference efficiency to some extent, they still suffer from issues like coarse-grained deployment and poor inference accuracy.Based on the discovery that different operators exhibit varying degrees of GPU affinity, an OATO(operator-aware tensor offloading)approach was proposed.OATO could extract operators′semantic knowledge and used it to design an intelligent scheduling algorithm, which further yielded a globally optimal model-deployment plan.Meanwhile, the OATO approach was integrated into the latest large model inference framework Llama.cpp to implement an operator-aware tensor offloading enhanced inference engine, referred to as OALlama.cpp.Experimental results show that compared with the state-of-the-art inference engines Llama.cpp and FlexGen, OALlama.cpp achieves the best inference performance on three large models.Notably, in the scenario where 75% of the LlaMA3-8B model weights are loaded on the GPU, the first-token generation speed of OALlama.cpp is nearly doubled compared with FlexGen and Llama.cpp.

AI chips face on-chip memory limits in deep learning.Current optimization methods focus on static computation graphs, leaving room to improve memory efficiency for dynamic graphs.To overcome this limitation, a memory optimization framework for control-flow computation graphs was developed.The framework realized operator-level memory reuse within subgraphs and further achieved recursive reuse across subgraphs by exploiting control-flow characteristics.In addition, a ping-pong buffering strategy for weight data was introduced to mitigate the memory wall between on-chip and off-chip memory, thereby allowing overlapping of memory access and computation operations within subgraphs.Validation on the domestic LUNA AI chip has demonstrated that the proposed framework improves on-chip memory utilization by 5.9% compared with existing methods.Moreover, the strategy effectively alleviates the memory wall problem by reducing data transfer time between on-chip and off-chip memory, resulting in execution efficiency improvements of up to 29%.

To achieve accurate and stable online identification of inductance parameters for PMSM(permanent magnet synchronous motor), an online inductance observation method based on virtual voltage vector excitation and current differential response was proposed, which required no additional test signal injection and was decoupled from rotor position, stator resistance, and permanent magnet flux linkage.By introducing the concept of a virtual voltage vector-oriented coordinate system, it was analytically derived and proven that the d-and q-axis inductances of a PMSM can be observed independently of the angular position in the conventional d-q synchronous reference frame.Building on this, the implementation procedure for extracting virtual voltage vectors and current differential information was discussed in detail, enabling non-intrusive inductance identification without any signal injection.The effectiveness and accuracy of the proposed method were validated by comparison with offline test procedures in IEEE standards.

For the common stator winding short circuit and rotor eccentricity faults in surface-mounted permanent magnet synchronous motors, a flexible printed circuit board with small footprint and capable of accommodating a large number of windings was used to fabricate the detection coil, which was then arranged in the stator slots to capture magnetic field information.For the stator winding short circuit fault, a winding short circuit detection method using dual orthogonal phase-locked loop to extract fault characteristic values was proposed.This method can effectively distinguish the short circuit resistance, short circuit winding number, and fault location, and was not affected by the motor′s speed fluctuations.For the rotor eccentricity fault, a differential bridge structure of the detection coil based on high-frequency injection was proposed for eccentricity detection, and ultimately, a 2% eccentricity detection can be achieved.For the composite fault, a fault discrimination scheme based on convolutional neural networks was introduced, and the performance of different learning methods was compared.The experimental results show that under the composite fault condition, a 98% correct rate of winding short circuit assessment is achieved, and the eccentricity detection error using AlexNet with a training data proportion of 60% is only 5%.

To address the challenges of intelligent diagnosis for open-circuit faults in power electronic inverters, such as the lack of actual fault samples and the issue of varying characteristic adaptability, a set of optimization methods was proposed from two key intelligent elements:data and algorithm, to support the practical applications of intelligent diagnosis for open-circuit faults in power electronic inverters.For the data element, a fault sample amplification method based on inverters′characteristics was proposed, which finds out the minimum number of practical samples required for model training.For the algorithm element, an attention-enhanced method and a frequency points adaptive training method for the diagnosis model were proposed, which significantly improve model training effectiveness and diagnosis accuracy under wide-frequency inverter operation.The effectiveness of the proposed optimization methods for the intelligent elements was validated by experiments.