@cluhmann sorry to interfere here, but unless they have put some new vodoo into the “more modern” MC algorithms, Metropolis should actually be superior in terms of speed (given close enough priors). You are right, it is certainly not a good choice, but the question was for speed.
@jsh980210 in my experience slow sampling is an indication of an over-complex or even wrong model design. I remember having the issue when defining model components which would contribute to the same slope; the model would randomly distribute values between them and never converge.
So in short, your general question can not be generally answered. Please provide an example (and, for me at least, the process of creating a minimal example is most instructive).
Also, sampling speed depends on your hardware and software. So you might consider sharing vague/approximate details about your setup so that the devs can exclude that as a bottleneck.
Cheers!