I/O is and will be limiting. That's probably why you see only a handful cores actually working.
Our fastq files range from 2G to 50G per read .. it sometimes takes ages to read the files (indepedently from tool).

same issue. I am using a cluster with client servers that each have 48 cores with hyperthreading up to 192 and want to trim PE fastq files that have ~ 400 million reads. I am reading from a dedicated raid 0 striped server which mitigates the I/O issue, but this tool ends up taking way longer than Trimmomatic, simply due to the time that it is taking to read the files and limiting its use to 16 threads.

I just hit this too using a 96 core machine on AWS. It would be nice if it were a warning but then attempted to use all the cores.

So is this limited to 16 threads or not? Because I also get this warning

Same issue..Why limit it to 16 threads only?

Since using >16 threads cannot make fastp faster. The bottleneck is I/O when thread number > 8.