You
should believe it, because it's true.
You were measuring malloc, so of course you came up with numbers that were 20 times worse than PyPy's nursery allocator. That's because malloc is 20 times slower, whatever common sense may say.
Also, you're talking about tail latency, while CF Bolz was measuring throughput. Contrary to your assertion, throughput is indeed a meaningful metric, though, especially for interactive UIs such as videogames, tail latency is often more important. For applications like compilers and SMT solvers, on the other hand, throughput matters more.
You're right that there are some other costs. A generational or incremental GC more or less requires write barriers, which slow down the parts of your code that mutate existing objects instead of allocating new ones. And a generational collector can move existing objects to a different address, which complicates FFIs that might try to retain pointers to them outside the GC's purview.
There are a lot of papers running captured allocation traces like you did against different memory allocators, some of them including GCed allocators. Ben Zorn did a lot of work like this in the 90s, coming to the conclusion that GC was somewhat faster than malloc/free. Both GCs and mallocs have improved substantially since then, but it would be easy to imagine that the results would be similar. Maybe someone has done it.
But to a significant extent we write performance-critical code according to the performance characteristics of our platforms. If mutation is fast but allocation is slow, you tend to write code that does less allocation and more mutation than you would if mutation was slow and allocation was fast, all else being equal. So you'd expect code written for platforms with fast allocation to allocate more, which makes faster allocation a bigger advantage for that code. Benchmarking captured allocation traces won't capture that effect.
Tail latency has always been one of the worst things about GCs. I hear amazing things about new real-time GC algorithms, including remarkable reports from Azul, but I don't know if they're true or what their price is.