In absolute terms, the numbers given in the PEP work out to about a thousand machine instructions per byte of input, for both the old and new parsers. From my point of view, that's pretty slow.

Note that this includes the time for tokenization, which is generally the bulk of the time spent in parsing, and which I believe did not change between the old recursive-descent parser and the new PEG parser. Typically the tokenizer must iterate over some 32 characters per line of input code, which it reduces to some 8 tokens for the benefit of the parser proper, so usually the tokenizer is the bottleneck. However, the performance in this case seems too poor to attribute to the tokenizer.

That PEP also discusses some avoidable inefficiencies in the old parser that accounted for a substantial fraction of its runtime.

One of the biggest advantages of PEGs is that you don't need a separate tokenization pass. But that means you are incurring Packrat's inefficiency per character instead of per token. As I said above, I believe the CPython implementation does not do this.

I've written and worked on PEG parser generators, and I think PEG parsing is the best option in a wide variety of circumstances. But it is rarely the fastest option and usually far from it.