Running finite element models with MSC NASTRAN, basically heavy matrix math. Matrices were NxN, N was around 10 million. This was on a server with 36 cores and a half terabyte of RAM, purchased in 2014.
Also seen Erlang workloads where you could get a bit of throughput increase with your VM scheduler scheduling more threads than your physical cores (so starting to use HT) but the latency would spike and become very unpredictable, which was a bad tradeoff for the use case.
HPC workloads normally at least won't benefit and probably take a hit from HT on Xeon-ish hardware at least. It's normally turned off on HPC compute nodes (perhaps in software so the resource manager can enable per-job if necessary). There are exceptions, particularly with KNC and, perhaps, KNL. The situation is likely different for POWER, but I don't have experience of it.
Also seen Erlang workloads where you could get a bit of throughput increase with your VM scheduler scheduling more threads than your physical cores (so starting to use HT) but the latency would spike and become very unpredictable, which was a bad tradeoff for the use case.