3 ms·
That's a great example - because they specifically built an Intel-compatible TSO memory model into the M1 CPUs that can be toggled at runtime to avoid a ton of
by arcticbull 5y ago
That's a great example - because they specifically built an Intel-compatible TSO memory model into the M1 CPUs that can be toggled at runtime to avoid a ton of fencing (and at about 70% the performance of ARM64 natively). This let Apple's Intel emulation on ARM run way faster than Microsoft's. They also added hardware support for ObjC/Swift reference counting. These are things no other vendor on earth would do.
- astrange 5y ago> They also added hardware support for ObjC/Swift reference counting. Not true. It just has fast atomics because it’s a single SoC.
- arcticbull 5y agoAh yes, it looks like it's just super-fast/almost-free uncontended atomic operations.
- pjmlp 5y agoChip support for reference counting goes back to 1960's microcoded CPUs, plenty of other vendors have done it.
- arcticbull 5y agoIndeed, and I was also wrong about that, to be clear. I did some research based on your comment, it's really cool too see. I'm a huge fan of old CPU architecture.
- captainmuon 5y ago> These are things no other vendor on earth would do I really wonder why. Microsoft has some great engineers. I would assume some of the compiler or kernel people regularly meet up with people from Intel or Qualcomm and talk about future development. It's not really the same, but when I was working in particle physics I remember some Intel engineers really trying hard to get into contact with researchers, and trying to figure out what we need, because they wanted to sell their Xeon Phi / Knights Landing cards to us (basically GPUs for general purpose computing based on X86, now discontinued).