I would make the point that C++ unordered_map is not remotely optimized for this use case, so it is not surprising that it is less efficient. You normally would not use it for something like this. In fact, C++ unordered_map is not tunable enough to offer excellent performance for most use cases. It is a very generic implementation of an unordered_map.

In any code base where performance matters, most important use cases for unordered_map would be replaced with a purpose-built implementation (like with GraphEngine). Not only is it pretty trivial to design and implement custom maps for various use cases, you can usually ensure at least a 2X improvement across all relevant metrics relative to the standard library unordered_map.

Based on the relative metrics for GraphEngine and C++ unordered_map in this article, I would expect a purpose-built C++ implementation to be significantly faster than GraphEngine. I know that is not what is being measured but it in real-world C++ implementations, it is how most properly engineered architectures would do it.