Test Metal.jl on ARM M-serie GPUs.
Mostly dev sandbox to try out new Metal.jl features on Apple's M2 processors.
Float32
is being supported. For Float64
, one could try using a construct from DoubleFloats.jl which may impact performance though.
Running diffusion_2D_kp.jl
on a MacBookAir with M2 chip results in
Perf. memcopy: time (s) = 0.00929, T_eff (GB/s) = 86.7
Perf. Laplace: time (s) = 0.00974, T_eff (GB/s) = 82.72 (0.95% of memcopy)
Recent blog posts with features highlights:
Gurrent version (29.03.23) requires Metal#main
.