BLAS call overhead when running multi-threaded

For julia 1.7 (and 1.6.4) running on Rosetta on m1, the default BLAS.get_num_threads() is 8 (for m1 pro, or 4 for m1); some operations are noticeably slower than running single-threaded, example:

using BenchmarkTools, LinearAlgebra


function make_lu(n)
    m = zeros(n, n)
    for i = 1:n
        v = rand(n)
        m .+= (0.1 + rand()^2) .* (v * v')
    end
    Hermitian(m)
end


println(BLAS.get_num_threads())
@benchmark lu!(a) setup = (a = make_lu(100))

Default (8 threads on m1 pro)

julia> @benchmark lu(a) setup = (a = make_lu(100))
BenchmarkTools.Trial: 1973 samples with 1 evaluation.
 Range (min … max):  149.041 μs … 153.483 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):       1.199 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):     1.684 ms ±   3.857 ms  ┊ GC (mean ± σ):  0.06% ± 1.45%

While

julia> BLAS.set_num_threads(1)

julia> @benchmark lu(a) setup = (a = make_lu(100))
BenchmarkTools.Trial: 6048 samples with 1 evaluation.
 Range (min … max):  67.375 μs … 309.958 μs  ┊ GC (min … max): 0.00% … 76.96%
 Time  (median):     69.417 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   70.368 μs ±   8.857 μs  ┊ GC (mean ± σ):  0.46% ±  2.89%

Is it some issue with openblas, or something wth julia?

julia> versioninfo()
Julia Version 1.7.0
Commit 3bf9d17731 (2021-11-30 12:12 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin19.5.0)
  CPU: Apple M1 Pro
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, westmere)
Environment:
  JULIA_MATHLINK = /Applications/Mathematica.app/Contents/Frameworks/mathlink.framework
  JULIA_MATHKERNEL = /Applications/Mathematica.app/Contents/MacOS/MathKernel

No repository file or test entry point is mentioned. Start by reproducing the supplied Julia benchmark on the stated Apple M1/Rosetta setup, comparing the default BLAS thread count with one thread. Done means identifying whether the overhead comes from BLAS/OpenBLAS or Julia and recording a confirmed explanation or actionable direction.