Introduction Neulite is a lightweight simulator designed for biophysically detailed neuron and network models. It consists of a frontend module, Bionetlite, and a numerical kernel. While the original kernel was optimized for Japan's flagship Supercomputer Fugaku, a CPU-based system, most contemporary supercomputers use CPU/GPU hybrid architectures. To adapt to these environments, we have developed a new kernel specifically for GPUs.
Methods We implemented Dendritic Hierarchical Scheduling (DHS), a GPU-based scheduling algorithm designed to solve linear equations on tree structures, such as dendritic trees [2]. To evaluate its performance, we utilized a balanced random network model, varying both the number of neurons and the threads per neuron. The excitatory-to-inhibitory ratio was maintained at 4:1. Benchmarks were conducted on a desktop system equipped with an Intel Core i5-14600KF CPU and an NVIDIA GeForce RTX 3070 Ti GPU.
Results Initially, we evaluated the effect of thread count per neuron by fixing the total number of neurons at 8,192. Under this condition, the optimal performance was achieved with 16 threads per neuron. Subsequently, using this optimal thread count, we varied the number of neurons to assess scalability. The results demonstrated that the GPU-based version consistently outperformed the CPU version as the network size increased. Specifically, at a scale of 8,192 neurons, the GPU implementation achieved a 16-fold speedup compared to the CPU version.
Discussion These results suggest that DHS is an efficient algorithm for simulating biophysically detailed neuron models, offering significant performance gains on GPU architectures. As a next step, we plan to evaluate the benchmark performance on more complex and biologically realistic network structures, such as cortical column models.
References
Acknowledgement Part of this study was supported by AMED Brain/MINDS 2.0 (JP25wm0625406). We would like to thank Drs. Kaaya Akira and Rin Kuriyama for technical discussions.
