IIRC they're a bit better than an Intel core on a single threaded load with some variance depending on the workload. But going from 1 to 2 threads per core increases the Power 9's throughput by 80% instead of the 20% you expect with an Intel core. And you can go up to 8 threads per core for something like 120% the total throughput of 1 thread. All those numbers for threads per core are also highly workload dependent. The more cache misses the more SMT helps.

Can you give a source, or was this personal testing?