4 ms·

Raspberry Pi OS 32-bit vs. 64-bit Performance

Havoc 5y ago
Wow. Much bigger differences than I was expecting. Anybody know how software support is looking for 64Bit? OS support is good but presumably the software itself needs to be arm 64 compiled too?
- makeworld 5y ago
  Since packages are derived from Debian (which supports arm 64), I assume pretty much all packages you can install from the Raspbian repos for 32 bit will still be available.
  - StillBored 5y ago
    Which begs the question of what is even the point of having their own distro anymore. Why not just push a bit more to land their changes upstream in linux, and make debian the default noobs install, while providing the option to do all the other OS's as well. I've said this before, that instead of focusing on building a custom distro everyone would benefit if they instead focused on just enabling all the other distro's and OS's. Pretty much everything, including windows runs on the darn thing, its the closest piece of HW in the arm ecosystem (other than the server class machines) to a random PC motherboard when it comes to just about working with any given OS.
    - yjftsjthsd-h 5y ago
      There's sense in trying to keep their diff from upstream fairly minimal, but I can see value in having some differentiation - if nothing else, the RPi foundation probably wants to ship Pi-branded images (default background, application menu icon, that kind of thing). Also, there might be issues around nonfree requirements, although I'm having a surprisingly hard time finding information about that; I know booting the Pi requires a nonfree firmware blob, but I'm not sure whether there are any ex. wifi driver issues. That said, yes one would very much hope that everyone can work together to make the diff between Debian and its downstreams as small as possible given the constraints that each party operates under.
  - Havoc 5y ago
    I see. That sounds very promising!
- pwinwood 5y ago
  Raspberry PI OS 32bit is compiled for ARMv6 to support the orginal Raspberry pi. I wonder what impact on benchmarks there would be compiling for ARMv7 hence only supporting Raspberry pi 2 and later?
- iam-TJ 5y ago
  I've been operating 64-bit kernels and user-space for at least a couple of years for embedded and network and server usage and not experienced anything to even remind me it's 64-bit rather than 32-bit. I've used both the original Raspbian 64-bit builds and plain Debian 64-bit with the few required RaspberryPi specific tools added in. Kernels are all self-built (cross-compiled) keeping up with mainline Linux.
matthewmacleod 5y ago
TLDR: Across the few dozen different workloads tested, switching Raspberry Pi OS 11 for the 64-bit version improved the performance on average by about 48% That's quite impressive!
em3rgent0rdr 5y ago
I suspect most of the performance improvements came not from being "64-bit", but rather from the new instructions in aarch64 (and even in armv7) over armv6 (which is what the 32-bit Raspbian has to use to be backwards compatible with the first Raspberry). So I don't think it is entirely appropriate to frame the performance improvements due to it going from "32-bit" to "64-bit". In fact, hypothetically assuming the same instruction set, 64-bit can cause performance degradation over 32-bit, particularly in memory-bound benchmarks, because the size of memory pointers and default int size will take up twice as much space, which puts stress on the limited cache space, which will generate additional off chip memory accesses.
- spacechild1 5y ago
  > default int size 'int' size stays the same. 'long' gets bigger (on LP64 platforms). Apart from that, I would agree with your assessment.
  - genpfault 5y ago
    https://en.wikipedia.org/wiki/64-bit_computing#64-bit_data_models https://en.wikipedia.org/wiki/64-bit_computing#64-bit_data_m...
    - em3rgent0rdr 5y ago
      wow thanks...learned something new. I'm sorry when I was in university there were still SPARC64 machines around, which use the "ILP64" where the default int size increased to 64 bits, so that must have been why I was taught that the default int size increased when going to 64-bit arch.
      - fulafel 5y ago
        Sun-shipped 64-bit Solaris has 32 bit int and it's in the LP64 column of the Wikipedia table. The entry of a ILP64 HAL systems version is about this other niche Solaris port for HAL designed chips that was not commonly used in universities (or anywhere, for that matter).
diimdeep 5y ago
> Using a 64-bit kernel (arm_64bit=1) just takes away 50MB of RAM, the worst choice is to combine 64-bit kernel and 64-bit userland since everything relevant get slower, same tasks require much more memory and the device will start to swap if low on memory or even kill processes due to out of memory. > As a rule of thumb you need almost twice as much RAM with a 64-bit userland compared to 32-bit with the same programs/services running. [1] [1]: https://github.com/ThomasKaiser/sbc-bench/blob/master/Results.md https://github.com/ThomasKaiser/sbc-bench/blob/master/Result...
- viraptor 5y ago
  That's not supported by any data posted there. You'd have to have almost all of your memory filled will pointers to achieve 2x increase which just doesn't happen in reality. Looking at a large application like Firefox https://www.raymond.cc/blog/mozilla-firefox-64bit-build-performance-compared-to-32bit/ https://www.raymond.cc/blog/mozilla-firefox-64bit-build-perf... you get ~30% increase.
- yjftsjthsd-h 5y ago
  >> As a rule of thumb you need almost twice as much RAM with a 64-bit userland compared to 32-bit with the same programs/services running. That doesn't make sense; going 64-bit should only double the size of pointers, it shouldn't touch the size of code or data in memory.

Wow. Much bigger differences than I was expecting. Anybody know how software support is looking for 64Bit? OS support is good but presumably the software itself needs to be arm 64 compiled too?

makeworld 5y ago
Since packages are derived from Debian (which supports arm 64), I assume pretty much all packages you can install from the Raspbian repos for 32 bit will still be available.
- StillBored 5y ago
  Which begs the question of what is even the point of having their own distro anymore. Why not just push a bit more to land their changes upstream in linux, and make debian the default noobs install, while providing the option to do all the other OS's as well. I've said this before, that instead of focusing on building a custom distro everyone would benefit if they instead focused on just enabling all the other distro's and OS's. Pretty much everything, including windows runs on the darn thing, its the closest piece of HW in the arm ecosystem (other than the server class machines) to a random PC motherboard when it comes to just about working with any given OS.
  - yjftsjthsd-h 5y ago
    There's sense in trying to keep their diff from upstream fairly minimal, but I can see value in having some differentiation - if nothing else, the RPi foundation probably wants to ship Pi-branded images (default background, application menu icon, that kind of thing). Also, there might be issues around nonfree requirements, although I'm having a surprisingly hard time finding information about that; I know booting the Pi requires a nonfree firmware blob, but I'm not sure whether there are any ex. wifi driver issues. That said, yes one would very much hope that everyone can work together to make the diff between Debian and its downstreams as small as possible given the constraints that each party operates under.
- Havoc 5y ago
  I see. That sounds very promising!
pwinwood 5y ago
Raspberry PI OS 32bit is compiled for ARMv6 to support the orginal Raspberry pi. I wonder what impact on benchmarks there would be compiling for ARMv7 hence only supporting Raspberry pi 2 and later?
iam-TJ 5y ago
I've been operating 64-bit kernels and user-space for at least a couple of years for embedded and network and server usage and not experienced anything to even remind me it's 64-bit rather than 32-bit. I've used both the original Raspbian 64-bit builds and plain Debian 64-bit with the few required RaspberryPi specific tools added in. Kernels are all self-built (cross-compiled) keeping up with mainline Linux.

matthewmacleod 5y ago

TLDR: Across the few dozen different workloads tested, switching Raspberry Pi OS 11 for the 64-bit version improved the performance on average by about 48% That's quite impressive!

em3rgent0rdr 5y ago

I suspect most of the performance improvements came not from being "64-bit", but rather from the new instructions in aarch64 (and even in armv7) over armv6 (which is what the 32-bit Raspbian has to use to be backwards compatible with the first Raspberry). So I don't think it is entirely appropriate to frame the performance improvements due to it going from "32-bit" to "64-bit". In fact, hypothetically assuming the same instruction set, 64-bit can cause performance degradation over 32-bit, particularly in memory-bound benchmarks, because the size of memory pointers and default int size will take up twice as much space, which puts stress on the limited cache space, which will generate additional off chip memory accesses.

spacechild1 5y ago
> default int size 'int' size stays the same. 'long' gets bigger (on LP64 platforms). Apart from that, I would agree with your assessment.
- genpfault 5y ago
  https://en.wikipedia.org/wiki/64-bit_computing#64-bit_data_models https://en.wikipedia.org/wiki/64-bit_computing#64-bit_data_m...
  - em3rgent0rdr 5y ago
    wow thanks...learned something new. I'm sorry when I was in university there were still SPARC64 machines around, which use the "ILP64" where the default int size increased to 64 bits, so that must have been why I was taught that the default int size increased when going to 64-bit arch.
    - fulafel 5y ago
      Sun-shipped 64-bit Solaris has 32 bit int and it's in the LP64 column of the Wikipedia table. The entry of a ILP64 HAL systems version is about this other niche Solaris port for HAL designed chips that was not commonly used in universities (or anywhere, for that matter).

diimdeep 5y ago

> Using a 64-bit kernel (arm_64bit=1) just takes away 50MB of RAM, the worst choice is to combine 64-bit kernel and 64-bit userland since everything relevant get slower, same tasks require much more memory and the device will start to swap if low on memory or even kill processes due to out of memory. > As a rule of thumb you need almost twice as much RAM with a 64-bit userland compared to 32-bit with the same programs/services running. [1] [1]: https://github.com/ThomasKaiser/sbc-bench/blob/master/Results.md https://github.com/ThomasKaiser/sbc-bench/blob/master/Result...

viraptor 5y ago
That's not supported by any data posted there. You'd have to have almost all of your memory filled will pointers to achieve 2x increase which just doesn't happen in reality. Looking at a large application like Firefox https://www.raymond.cc/blog/mozilla-firefox-64bit-build-performance-compared-to-32bit/ https://www.raymond.cc/blog/mozilla-firefox-64bit-build-perf... you get ~30% increase.
yjftsjthsd-h 5y ago
>> As a rule of thumb you need almost twice as much RAM with a 64-bit userland compared to 32-bit with the same programs/services running. That doesn't make sense; going 64-bit should only double the size of pointers, it shouldn't touch the size of code or data in memory.