In simulation of nuclear reactor physics using the Monte Carlo neutron transport method on GPUs, the sorting of particles play a significant role in execution performance. Traditionally, CPUs and GPUs are separated devices connected with low data transfer rate and high data transfer latency. Emerging computing chips tend to integrate CPUs and GPUs. One example is the Apple silicon chips with unified memory. Such a unified memory chips has opened doors for new strategies of collaboration between CPUs and GPUs for Monte Carlo neutron transport. Sorting particle on CPU and transport on GPU is an example of such new strategy, which has been suffering the high CPU-GPU data transfer latency on the traditional devices with separated CPU and GPU. The finding is that for the Apple M2 max chip, sorting on CPU leads to better performance than sorting on GPU for the ExaSMR whole core benchmark problems, while for the HTR-10 high temperature gas reactor fuel pebble problem, sorting on GPU is more efficient. The features of partially sorted particle order have been identified to contribute to the higher performance with CPU sort than GPU for the ExaSMR problem. The in-house code using both CPUs and GPUs achieves 7.5 times power efficiency that of OpenMC on CPUs for ExaSMR whole core and 50 times for HTR-10 fuel pebble benchmark problems.

翻译：在利用GPU进行反应堆蒙特卡洛中子输运的核反应堆物理模拟中，粒子的排序对执行性能起着重要作用。传统上，CPU和GPU是分离的设备，通过低数据传输速率和高数据传输延迟的接口相连。新兴的计算芯片倾向于将CPU和GPU集成在一起，例如采用统一内存的Apple硅芯片。这种统一内存芯片为蒙特卡洛中子输运中CPU与GPU之间新的协作策略打开了大门。在CPU上进行粒子排序并在GPU上执行输运便是这种新策略的一个实例，传统上该策略因分离式CPU-GPU架构下的高数据传输延迟而受到限制。研究发现，对于Apple M2 max芯片，在ExaSMR全堆芯基准问题中，CPU排序的性能优于GPU排序；而对于HTR-10高温气冷堆燃料球问题，GPU排序更具效率。部分有序的粒子排列特性被确定为ExaSMR问题中CPU排序比GPU排序性能更高的关键因素。自主研发的CPU-GPU协同计算代码在ExaSMR全堆芯和HTR-10燃料球基准问题中，能效分别达到OpenMC在CPU上运行的7.5倍和50倍。