How does Architecture Influence the Base Capabilities of Pre-trained Language Models? A Case Study Based on Wider Transformer Models