greta benchmarks

Each post compares versions of greta. A post’s run.R holds all of the code that was measured: the models, the calls that were timed, and the settings. It installs each version with {cross} and times it with {bench}, in a fresh R session per version, and the page shows that script in full, beside every result it produced.

Date	Title	Description
Oct 13, 2026	greta#855 (tf-warmup-i547): effective samples from its warmup tuning	Effective samples per sampler iteration and per second of `mcmc()` with each version’s defaults on main and greta#855, on four of greta’s examples over twenty sessions. #855’s warmup tunes to larger step sizes, gives more effective samples per sampler iteration on three of the four models, and gives 2.8 to 7.9 times main’s effective samples per second.
Oct 7, 2026	greta#855 (tf-warmup-i547): call overhead in `mcmc()` and `opt()`	Runtime per call of `mcmc()` and `opt()` on greta#855 with and without its last two changes, which keep `opt()`’s traced loop and `mcmc()`’s chain state in TensorFlow between calls. Without them, a call to `opt()`, or to `mcmc()` with `one_by_one = TRUE`, takes several times as long.
Oct 7, 2026	greta#855 (tf-warmup-i547): warmup and sampling in TensorFlow	How long mcmc() and opt() with adam() take on greta 0.6.0 from CRAN, main, greta#850 and greta#855, on four of greta’s example models. greta#855 is the fastest of the four at both, and its mcmc() samples the same posteriors.
Oct 7, 2026	greta#855 (tf-warmup-i547): TFP’s loops, tried and not adopted	TensorFlow Probability’s sample_chain() and minimize() timed against greta’s own sampling loop and opt(), on the PR’s branch and two models. Once traced, sample_chain() takes as long as greta’s loop, and minimize() takes longer than opt(), so the PR keeps greta’s own loops.
Oct 7, 2026	greta#853 (windowed-adaptation-i853): warmup schemes compared	Five warmup schemes run inside TensorFlow on five models: greta’s own, greta’s with greta#841’s count fixed, Stan’s windowed scheme from greta#853, and TFP’s windowed HMC and NUTS. Stan’s scheme finds the scales of the normals that greta’s own misses, and gives more effective samples per second than greta’s on every model but eight_schools.
Oct 6, 2026	greta#833 (wire-jit-compile): XLA compilation	Building a model, opt() and mcmc() on greta’s CRAN release, main and greta#833, all with model(compile = TRUE), which only #833 acts on. With XLA, opt() is slower, building a model and mcmc() take about as long, and the posteriors are the same.
Oct 5, 2026	greta#843 (faster-hessians-i546): retracing and hessians	Building models, opt(), opt(hessian = TRUE) and mcmc() on greta 0.6.0 from CRAN, on main and on greta#843. greta#843 traces each model’s TensorFlow functions fewer times, and opt(hessian = TRUE) is much faster on it.