undefined | Better HN

0 pointsricardobeat1y ago0 comments

Repeating what I just commented elsewhere, but Mac uses several advanced memory management features: apps can share read-only memory for common frameworks, it will compress memory instead of paging out, better memory allocation, less fragmentation.

Bandwidth for copying things into memory is also vastly faster than what you get on Intel/AMD, for example on the Max chips you get 800GB/s which is the rough equivalent of 16 channels of DDR5-6400, something simply not available in consumer hardware. You can get 8 channels with AMD Epyc, but the motherboard for that alone will cost more than a Mac mini.

0 comments

Rohansi1y ago

Sharing read-only/executable memory and compressed memory are also done on Windows 10+ and modern Linux distributions. No idea what "better memory allocation" and "less fragmentation" are.

800GB/s is a theoretical maximum but you wouldn't be able to use all of it from the CPU or even the GPU.

https://www.anandtech.com/show/17024/apple-m1-max-performanc...

hollerith1y ago

>apps can share read-only memory for common frameworks

How is that different from plain shared libraries?

ricardobeatOP1y ago

System design and stability. On MacOS a lot is shared between applications compared to the average Linux app. Dynamic linking has fallen out of favor in Linux recently [1], and the fragmentation in the ecosystem means apps have to deal with different GUI libraries, system lib versions etc, whereas on Mac you can safely target a minimum OS version when using system frameworks. Apps will also rarely use third party replacements as the provides libraries cover everything [2], from audio to image manipulation and ML.

[1] https://lore.kernel.org/lkml/CAHk-=whs8QZf3YnifdLv57+FhBi5_W...

[2] https://developer.apple.com/library/archive/documentation/Ma...

The_Colonel1y ago

People who need 64GB+ RAM are not running 1000 instances of native Apple apps. They run docker, VMs, they run AI models, compile huge projects, they run demanding graphics applications or IntelliJ on huge projects. Rich system libraries are irrelevant in these cases.

1 more reply

jcelerier1y ago

I can run apps with 4 distinct toolkits on Linux and memory usage will barely go past the memory usage of opening one Facebook or Instagram tab in a browser. Compared to compiling a single semi-large source file with -fsanitize=addresses which can cause one single instance of GCC or Clang to easily go past 5G of memory usage no matter the operating system...

j / k navigate · click thread line to collapse