I recently obtained access to a newer Mac computer with ARM processor. Is there some ready to use benchmark I can execute to see if running native ARM Julia will be faster (and safe) vs running Julia x86 on ARM through Rosetta 2?
You can use SystemBenchmark.jl.
I had already run a benchmark of a native build vs Rosetta at Submit benchmark files here · Issue #8 · IanButterworth/SystemBenchmark.jl · GitHub
Not what you asked, but at Submit benchmark files here · Issue #8 · IanButterworth/SystemBenchmark.jl · GitHub there is also a comparison Linux vs macOS on the same machine
From a totally non-quantitative perspective M1 ARM has been much much faster than Rosetta, so that I switched to using it even before it was stable (back with 1.7). On the 1.8 beta releases I have had no M1 specific issues at all arise.
I have segfaults on M1 when differentiate relatively large models with Zygote. Otherwise no issues since 1.7
Are there silent errors, or can I rule that out?
I haven’t come across any and my test sets are fairly extensive.