New computer...Mac or Windows, RStan or CmdStanR?

We don’t prioritize fixing Mac/Linux bugs. A lot of our users are on Windows.

If you do go with Windows, you can use the Linux Subsystem for Windows to run Stan.

Stan doesn’t link to external matrix libs, so this isn’t an issue. (It is possible to configure Eigen to use BLAS, but almost none of our users or devs do that.)

If you’re doing C++, the Linux and Mac versions tend to have more robust and more standard tooling around them than Windows.

RStan is what you need if you want to get log densities, gradients, and transforms in R. Otherwise, cmdstanr is more up to date with Stan.

For simple model compilation and fitting, they’re very similar in the way they work from a user perspective. They’re different under the hood, but the interfaces are similar.

The speed should be similar, but CmdStanR can work in less memory.

Take that as a good sign. People post when they can’t install something.

Pretty much all the reports are that the M1 Macs are very fast running Stan (and browsers and basically anything else that keeps lots of memory open). I’m hoping to get one of the M2 MacBook Pros soon and see how those work.

I’d love to see some benchmarks from people who have all this hardware at hand, even on simple problems.

1 Like

I have several different pieces of hardware available. Do you have a test in mind you’d like run?

Sure. I’d like to evaluate

  • Intel (2 or 4 or 8 real cores), M1, M2 hardware

using

  • 1, 2, 4, 8 parallel chains

for a problem where the data’s big enough to cause cache misses. Easiest thing to do would be to simulate a logistic regression of 100K data items with 100 predictors each. Then see how long it takes to run using our default settings.

2 Likes

I also have access to an M1 Mac Studio; if you want I can run some tests. I primarily run CmdStanR but have RStan installed as well.

Just point me to the model/data and let me know what you want me to do :-)