That example is amongst the worst synthetic benchmarks I have seen in a while. Also it comes from someone who is associated with piwik. A PHP project that makes wordpress look tidy, well engineered and speedy.
The more deeply nested the callstack is when you supress, the bigger the performance hit (slight generalisation and slightly from memory).
I spent 10-15 minutes poking around php internals newslist for an explanation of the php 5.4 changelog "improved performance of the (@) silence operator" without luck. It's just listed as a zend engine improvement without specifications.
Someone posted here a clear example of why (in earlier php5) it generates far worse opcode:
The more deeply nested the callstack is when you supress, the bigger the performance hit (slight generalisation and slightly from memory).