Statistical modelling front-ends that run on Apple GPU hardware via the Rmlx array library.
GPUs are designed to handle matrices, which is a good fit for statistics. But up till now R Mac users have not had access to the power of their GPUs. RmlxStats is an experiment in implementing common statistical methods on the GPU. RmlxStats is early work in progress!
Functions implemented so far include Rmlx versions of lm, glm, glmnet and a bootstrapping function mlxs_boot().
When to use
RmlxStats offers large speedups against both base R functions, and speed-optimized packages like speedglm and RCppEigen. Speedups are especially large for regressions with many predictors (large p).
Very roughly, if you have 50 or more predictors and 10,000 or more rows, or if your regressions are taking measurable time to complete, RmlxStats is worth trying:
# On my machine
> system.time({lm <- lm(arr_delay ~ dep_delay + factor(paste(month,day)), data = nycflights13::flights); })
user system elapsed
31.310 0.179 31.479
> system.time({lm2 <- mlxs_lm(arr_delay ~ dep_delay + factor(paste(month,day)), data = nycflights13::flights); Rmlx::mlx_eval(lm2$coefficients)})
user system elapsed
4.421 0.271 2.818 See the benchmarks vignette for more details.
GPU calculations use float32 precision, so if you need higher numerical accuracy than this, RmlxStats may not be the right tool.
Installation
Install Apple’s MLX runtime:
Then:
remotes::install_github("hughjonesd/RmlxStats")which will also install Rmlx.