Poor vDSP performance on M1

Anyone else noticing really bad performance of FloatVectorOperations and vDSP on arm macs?

If I just do some vDSP_vmul, the code runs 5 times faster on an intel mbp (2.6Ghz i7) than on a m1Pro mbp here (not using rosetta).

Am I doing something wrong or is that kind of expected? What’s your experience with that?


Would be interesting to also compare its performance to running in Rosetta on the same machine

1 Like

The thing you are maybe seeing is that your code in not CPU intensive enough hence is put on the efficiency core


Same here, with plugins using my own interface to IPP / vDSP.

I use IPP on Intel Macs, because vDSP is significantly slower than IPP for my plugins.

I cannot imagine that macOS is constantly redistributing executing code over its cores.

@otristan : would that be possible if your code only runs in the host’s callback and thus on the same RT thread as the callback?

I doubt it, but you can easily check which core is active on your system.