Within-chain parallelization misbehaves with brms in a model with measurement errors

Thanks @paul.buerkner, much clearer now.

I am not sure I follow your point 1.

What I am saying is that anything which has a high computational cost per observation is a good candidate for efficient parallelization.

Not easy to parallelize (examples):

  • normal density
  • bernoulli or bernoull logit

Much better to parallelize (these require the evaluation of the log-Gamma function):

  • Poisson
  • Negative Binomial
1 Like