That’s oversimplified. If the falsely predicted branch issues a load instruction for example, you might use up valuable memory bandwidth, trash cache lines, etc

That’s oversimplified. If the falsely predicted branch issues a load instruction for example, you might use up valuable memory bandwidth, trash cache lines, etc. There are way more side effects that can make false branch predictions more expensive. It’s also extremely microarchitecture dependent, proprietary and undocumented what exactly happens, which is why we need to benchmark the actual code under consideration.