The new M4 Apple Silicon chips look to have pretty impressive performance. Has anyone tried Stan on the an M4 Mac yet? If so, have you had any technical issues? Are you getting noteworthy speed ups compared to your old set-up?

The M2 and M3 Macs perform really well for Stan. This is because we’re largely memory bottlenecked in large models and the ARM series is much better at dealing with asynchronous memory access.

I got one of the base-level $500 M4 Mac Minis to play around with – have only run a few Stan models on it so far, but noticing a 1.5-3x speedup relative to my laptop (a 2019 MBP w/ an i9-9980HK), which is consistent with the ~2.5x difference in Geekbench 6 scores. Note also this was using 8 chains i…

@jroon @Bob_Carpenter this might also interest you – I did a quick test of within-chain parallelization via reduce_sum in a toy model, comparing my Macbook’s performance to the new M4 Mac Mini’s: functions { real partial_sum_lpdf(array[] real y_subset, int start, int end, real mu, real sigma) { …

Thanks for the analysis, @NikVetr . [image] NikVetr: there are benefits to within-chain parallelization in this implementation even with very small models (N=500) – I’d have expected overhead to dominate even at 2 threads. You see this with JAX models, too. It depends on whether your hardwar…

[image] NikVetr: with BF sales, I could get a base M4 Mac Mini for ~$400, so I snagged another 3 and currently have them daisy chained together and to my macbook via TB4 bridge, distributing jobs – including Stan runs – via GNU Parallel. So I can either distribute embarassingly parallelizable ru…

hm, not off the top of my head – the thunderbolt bridge stuff just entailed plugging them all in, going to System Settings > Network > Thunderbolt Bridge, and then Manually configuring them to have the same Subnet mask (I used 255.255.255.0) and different IP addresses (I used 10.0.0.1, 10.0.0.2, 10.…

Thanks very much thats gives me somewhere to start! In terms of the workflow from the Stan point of view - you are first compiling the Stan program on the main computer and then via the Shell distributing jobs to the worker machines? (i.e. I am assuming you cannot use this distributive computing ac…

That’s a great analysis! @NikVetr would you mind trying to run this using cmdstan’s across chain sampling? I’d be curious how much sharing the data would help the small and large data models get closer to the peak efficiency

Nice comparison! A grainsize of 1 is not ideal in this setting. I wonder if one could get better efficiency for the larger sample sizes if one would use some other heuristic here. I’d try a grain size relative to the number of cores and sample size, e.g. to have around 10 chunks of work per CPU. Th…

No problem! Feel free to ping me if you have any questions, too a bit of trial and error to get everything up and running. :] you are first compiling the Stan program on the main computer and then via the Shell distributing jobs to the worker machinesA ah so that’s one hitch I encountered – even…

Stan on M4 Mac?

General

WardBrian February 6, 2025, 8:34pm 16

We did:

https://mc-stan.org/docs/cmdstan-guide/parallelization.html#running

https://mc-stan.org/docs/cmdstan-guide/mcmc_config.html#sampler-num-chains

Topic		Replies	Views
Four chains vs four jobs General cmdstan	28	761	June 19, 2024
Using Stan on a computing cluster. Any advice? CmdStan	20	5474	January 10, 2019
Cmdstanpy, mpi speedup Developers	26	763	November 19, 2024
Hardware Advice General	21	3465	October 8, 2020
RStan (PyStan) & MPI / GPU Developers features	43	4033	September 24, 2017

Stan on M4 Mac?

Related topics