The 8x improvement is somewhat misleading as it is based on a single TPC-H query. The other queries from the benchmark seem to show a ~3x improvement: http://vitessedata.com/technology

Given that their query engine uses an LLVM JIT compared to the standard PostgreSQL interpreter, these results are very reasonable. This is similar to comparing an interpreter vs a JIT for a programming language.

Even with the best intentions, benchmarking a database is notoriously difficult. I work on Presto, a distributed analytic SQL database, and we are often asked how it compares to other systems, or how it will perform for a given workload. The answer is always "try it and see", both because how it performs for your configuration and workload is the only thing that matters, and because there are so many variables that factor into the result.