4 ms·

Arm64 Windows running with hardware virtualization on Apple Silicon

boromi 6y ago
why does arm64 need virtualization on apple silicon? Isn't M1 chip referenced on the Arm architecture?
- krackers 6y ago
  It's virtualization, not emulation. It is indeed native arm, which is why virtualization suffices. To run x64 windows you would need to emulate. The virtualization is needed because you are running the guest OS on top of the host OS. The same situation as if you were running x64 windows on an intel mac (e.g. via Virtualbox, vmware, or Parallels): in both cases you need kernel-level support in the host OS so that it can act as a hypervisor, and hardware support from the processor (in Intel it's done via vt-x) to support things like nested paging, ring -1, etc.
  - cordite 6y ago
    Is Microsoft planning on arm emulation of x64 as well? Seems like a very interesting future if they are.
    - tambourine_man 6y ago
      They already do: https://www.theverge.com/2020/9/30/21495510/microsoft-windows-on-arm-x64-app-emulation https://www.theverge.com/2020/9/30/21495510/microsoft-window...
    - rowanG077 6y ago
      They have been doing for more then a year.
  - ineedasername 6y ago
    My knowledge may be out of date here, but I seem to recall that with a quality hypervisor (ESXi and even VMware workstation running on a host OS) you can see performance around 90-95% of bare metal. Of course a type 2 hypervisor will depend on what the host OS is doing, but if it's working at a 50% load then the guest OS should perform almost as though it had 50% available. (Yes, lots of variables will factor in here) Not sure if we'll get there on an M1 without a heck of a lot more work, but it's early days and things show lots of promise. What I really want though is not to have to own Apple hardware for this. The M1 shows such promise that it would be crazy for this sort of chip not to be pursued as the next level of commodity chips.
    - Tagbert 6y ago
      A significant reason for the high performance of Mac OS on M1 is that the M1 has been tweaked to optimize common Mac OS/ObjectiveC/Swift operations. In one interview, Apple engineers said that they worked on speeding up the operation to allocate and deallocate objects by 5x over Intel chips. That kind of operation is common and used in loops and in the reference count garbage collection used by those languages. By doing this kind of optimization they can tune the SOC to the OS. That is hard to do for commodity chips.
      - fomine3 6y ago
        Could you point the interview?
        CyanLite4 6y ago
        https://twitter.com/Catfish_Man/status/1326238434235568128?s=20 https://twitter.com/Catfish_Man/status/1326238434235568128?s...
      - ineedasername 6y ago
        Given Windows on the consumer side & linux on the server side, is there any reason, (assuming difference chip design paths for server vs consumer) that similar optimizations couldn't become fairly standard on commodity chips? I'd think MS especially would be willing to have some tight partnerships with chip devs to do so.
seba_dos1 6y ago
Are we going to get a separate post for every single operating system and distribution that someone decided to try in a VM on the new Apple hardware?
- tambourine_man 6y ago
  Windows is kind of a big deal, being ≈90% of the desktop market and closed source. But sure, I'd love to have a post for Ubuntu, CentOS, Arch… Why, any objections?
  - tinus_hn 6y ago
    Actually there was a post about Ubuntu today: https://news.ycombinator.com/item?id=25219698 https://news.ycombinator.com/item?id=25219698
  - xiphias2 6y ago
    The next interesting thing for me is PyTorch support, as Tensorflow is already running. I still don't know what data types and operations the neural engine supports though for example.
- ineedasername 6y ago
  Here's the thing: I know it's a lot of M1 news, and I thought it a bit much myself. But then I realized just how much of a game changer this chip is. Not just for Apple, but because the M1 success will really spark and accelerate other efforts at more commodity chips with similar capabilities. The problem though is chip architecture and the consequent software compatibility issues, and when compatible the performance issues. So, given the potential future of this choice of architecture and chip design, every development in compatibility or performance really does have significant consequences. And of course, the recommended community guidelines for dealing with posts you don't think are relevant or newsworthy to HN is to flag them instead of complaining in comments. If enough people share your opinion, it will disappear. If not, then it is de facto relevant to the community.
  - 6y ago

why does arm64 need virtualization on apple silicon? Isn't M1 chip referenced on the Arm architecture?

krackers 6y ago
It's virtualization, not emulation. It is indeed native arm, which is why virtualization suffices. To run x64 windows you would need to emulate. The virtualization is needed because you are running the guest OS on top of the host OS. The same situation as if you were running x64 windows on an intel mac (e.g. via Virtualbox, vmware, or Parallels): in both cases you need kernel-level support in the host OS so that it can act as a hypervisor, and hardware support from the processor (in Intel it's done via vt-x) to support things like nested paging, ring -1, etc.
- cordite 6y ago
  Is Microsoft planning on arm emulation of x64 as well? Seems like a very interesting future if they are.
  - tambourine_man 6y ago
    They already do: https://www.theverge.com/2020/9/30/21495510/microsoft-windows-on-arm-x64-app-emulation https://www.theverge.com/2020/9/30/21495510/microsoft-window...
  - rowanG077 6y ago
    They have been doing for more then a year.
- ineedasername 6y ago
  My knowledge may be out of date here, but I seem to recall that with a quality hypervisor (ESXi and even VMware workstation running on a host OS) you can see performance around 90-95% of bare metal. Of course a type 2 hypervisor will depend on what the host OS is doing, but if it's working at a 50% load then the guest OS should perform almost as though it had 50% available. (Yes, lots of variables will factor in here) Not sure if we'll get there on an M1 without a heck of a lot more work, but it's early days and things show lots of promise. What I really want though is not to have to own Apple hardware for this. The M1 shows such promise that it would be crazy for this sort of chip not to be pursued as the next level of commodity chips.
  - Tagbert 6y ago
    A significant reason for the high performance of Mac OS on M1 is that the M1 has been tweaked to optimize common Mac OS/ObjectiveC/Swift operations. In one interview, Apple engineers said that they worked on speeding up the operation to allocate and deallocate objects by 5x over Intel chips. That kind of operation is common and used in loops and in the reference count garbage collection used by those languages. By doing this kind of optimization they can tune the SOC to the OS. That is hard to do for commodity chips.
    - fomine3 6y ago
      Could you point the interview?
      - CyanLite4 6y ago
        https://twitter.com/Catfish_Man/status/1326238434235568128?s=20 https://twitter.com/Catfish_Man/status/1326238434235568128?s...
    - ineedasername 6y ago
      Given Windows on the consumer side & linux on the server side, is there any reason, (assuming difference chip design paths for server vs consumer) that similar optimizations couldn't become fairly standard on commodity chips? I'd think MS especially would be willing to have some tight partnerships with chip devs to do so.

Are we going to get a separate post for every single operating system and distribution that someone decided to try in a VM on the new Apple hardware?

tambourine_man 6y ago
Windows is kind of a big deal, being ≈90% of the desktop market and closed source. But sure, I'd love to have a post for Ubuntu, CentOS, Arch… Why, any objections?
- tinus_hn 6y ago
  Actually there was a post about Ubuntu today: https://news.ycombinator.com/item?id=25219698 https://news.ycombinator.com/item?id=25219698
- xiphias2 6y ago
  The next interesting thing for me is PyTorch support, as Tensorflow is already running. I still don't know what data types and operations the neural engine supports though for example.
ineedasername 6y ago
Here's the thing: I know it's a lot of M1 news, and I thought it a bit much myself. But then I realized just how much of a game changer this chip is. Not just for Apple, but because the M1 success will really spark and accelerate other efforts at more commodity chips with similar capabilities. The problem though is chip architecture and the consequent software compatibility issues, and when compatible the performance issues. So, given the potential future of this choice of architecture and chip design, every development in compatibility or performance really does have significant consequences. And of course, the recommended community guidelines for dealing with posts you don't think are relevant or newsworthy to HN is to flag them instead of complaining in comments. If enough people share your opinion, it will disappear. If not, then it is de facto relevant to the community.
- 6y ago