Reconfigurable array architecture has become an important hardware platform for edge-side deployment of convolutional neural networks due to their high parallelism and flexible programmability.However,traditional mult...Reconfigurable array architecture has become an important hardware platform for edge-side deployment of convolutional neural networks due to their high parallelism and flexible programmability.However,traditional multi-branch convolutional networks suffer from computational redundancy,high memory access overhead,and inefficient branch fusion.Therefore,this paper proposes an adaptive multi-branch convolutional module(AMBC)that integrates software-hardware co-optimization.During training,the learnable fusion coefficients are introduced to enable adaptive fusion of multi-scale features,while in the inference phase,the multiple branches and their normalization parameters are merged with the fusion coefficients into a single 3×3 convolutional kernel through operator fusion.On the SIREA-288 reconfigurable platform,compared with unoptimized multi-branch networks,the proposed AMBC reduces external memory accesses by 47.91%and inference latency by 47.20%,achieving a 1.90×speedup.This approach maximizes the utilization of the reconfigurable logic while minimizing both reconfiguration and data-movement overheads in edge inference.展开更多
Fusion power output is proportional not only to the fuel particle number densities participating in reaction but also to the fusion reaction rate coefficient （or reactivity）, which is dependent on reactant velocity ...Fusion power output is proportional not only to the fuel particle number densities participating in reaction but also to the fusion reaction rate coefficient （or reactivity）, which is dependent on reactant velocity distribution functions. They are usuMly assumed to be dual Maxwellian distribution functions with the same temperature for thermal nuclear fusion circumstances. However, if high power neutral beam injection and minority ion species ICRF plasma heating, or multi-pinched plasma beam head-on collision, in a converging region are required and investigated in future large scale fusion reactors, then the fractions of the injected energetic fast ion tail resulting from ionization or charge exchange will be large enough and their contribution to the non-Maxwellian distribution functions is not negligible, hence to the fusion reaction rate coefficient or calculation of fusion power. In such cases, beam-target, and beam-beam reaction enhancement effect contributions should play very important roles. In this paper, several useful formulae to calculate the fusion reaction rate coefticient for different beam and target combination scenarios are derived in detail展开更多
基金Supported by the National Science and Technology Major Project of China(2022ZD0119005)the Natural Science Project of Shaanxi Province(2025JC-YBMS-754,2024JC-YBMS-539)。
文摘Reconfigurable array architecture has become an important hardware platform for edge-side deployment of convolutional neural networks due to their high parallelism and flexible programmability.However,traditional multi-branch convolutional networks suffer from computational redundancy,high memory access overhead,and inefficient branch fusion.Therefore,this paper proposes an adaptive multi-branch convolutional module(AMBC)that integrates software-hardware co-optimization.During training,the learnable fusion coefficients are introduced to enable adaptive fusion of multi-scale features,while in the inference phase,the multiple branches and their normalization parameters are merged with the fusion coefficients into a single 3×3 convolutional kernel through operator fusion.On the SIREA-288 reconfigurable platform,compared with unoptimized multi-branch networks,the proposed AMBC reduces external memory accesses by 47.91%and inference latency by 47.20%,achieving a 1.90×speedup.This approach maximizes the utilization of the reconfigurable logic while minimizing both reconfiguration and data-movement overheads in edge inference.
基金Supported by the International Thermonuclear Experimental Reactor Project of China under Grant No 2013GB114003the National Natural Science Foundation of China under Grant No 11275135
文摘Fusion power output is proportional not only to the fuel particle number densities participating in reaction but also to the fusion reaction rate coefficient （or reactivity）, which is dependent on reactant velocity distribution functions. They are usuMly assumed to be dual Maxwellian distribution functions with the same temperature for thermal nuclear fusion circumstances. However, if high power neutral beam injection and minority ion species ICRF plasma heating, or multi-pinched plasma beam head-on collision, in a converging region are required and investigated in future large scale fusion reactors, then the fractions of the injected energetic fast ion tail resulting from ionization or charge exchange will be large enough and their contribution to the non-Maxwellian distribution functions is not negligible, hence to the fusion reaction rate coefficient or calculation of fusion power. In such cases, beam-target, and beam-beam reaction enhancement effect contributions should play very important roles. In this paper, several useful formulae to calculate the fusion reaction rate coefticient for different beam and target combination scenarios are derived in detail
