Computational Bottlenecks of Training Small-scale Large Language Models