4 ms·

Ultimately Apple has hit the same dead end too. Their designs and TSMC's processes aren't putting Moore's law back on the map, you can see it in the year-over-y

Ultimately Apple has hit the same dead end too. Their designs and TSMC's processes aren't putting Moore's law back on the map, you can see it in the year-over-year performance. If tariffs end up hitting any harder, Apple can't just go to TSMC's Arizona plant and get faster silicon than what Taiwan has today. > they screwed up because they said Linux, which doesn’t exclude Apple Silicon In the context of the second graph, that means Apple Silicon is also failing to push the envelope forwards.

altairprime 2y ago
Apple has one advantage that PCs do not: each silicon release cycle is accompanied by a coordinated operating system and compiler release for CPU, GPU, and NPU workloads; so that they can hardware-accelerate functions inside the operating system and programs compiled to target the hardware far more rapidly than the PC industry. I remember a tweet from a couple years ago where the hardware-accelerated IPC calls for a system-wide performance gain (and thus a latency reduction slash benchmark improvement) of several percent within one of those year-long intervals — but such gains would be invisible to Linux, even Asahi, because it would require libc and kernel adaptations for different hardware generations, as well as reverse engineering those adaptations from the various macOS kernels and frameworks each year. I don’t think all of those gains can be measured relative to PCs at all, as there may not be any comparable equivalent of what’s being accelerated for PCs to measure — but it still stands out that they simply chose not to mention Apple at all, when they could have theoretically shown either the same slowdown or lack thereof in that domain now that multiple generations are available for easy testing.
- talldayo 2y ago
  That's neither a new concept, nor unique to Apple. GCC and LLVM are both regularly updated by AMD and Intel who also optimize and patch their own microcode as well. When you compile the Linux kernel, and moreover Linux software, you will be doing it with the optimizations of your chosen architecture. For x86 that means OEMs provide optimizations for stuff like AVX and SSE, whereas Apple works in their equivalent proprietary frameworks like Accelerate. You can credit Apple (or more accurately, Lattner) with starting the trend, but Linux can be compiled using the same LLVM backend MacOS uses with whatever optimizations your OEM supports. > I remember a tweet from a couple years ago where the hardware-accelerated IPC calls for a system-wide performance gain Most kernels aren't bottlenecked by IPC throughput like XNU is. A modern Linux system might need IPC for dbus, a few journaling systems and systemd. Apple's entire kernel relies on the throughput of a message queue, so feasibly speeding that up matters quite a lot.

altairprime 2y ago

Apple has one advantage that PCs do not: each silicon release cycle is accompanied by a coordinated operating system and compiler release for CPU, GPU, and NPU workloads; so that they can hardware-accelerate functions inside the operating system and programs compiled to target the hardware far more rapidly than the PC industry. I remember a tweet from a couple years ago where the hardware-accelerated IPC calls for a system-wide performance gain (and thus a latency reduction slash benchmark improvement) of several percent within one of those year-long intervals — but such gains would be invisible to Linux, even Asahi, because it would require libc and kernel adaptations for different hardware generations, as well as reverse engineering those adaptations from the various macOS kernels and frameworks each year. I don’t think all of those gains can be measured relative to PCs at all, as there may not be any comparable equivalent of what’s being accelerated for PCs to measure — but it still stands out that they simply chose not to mention Apple at all, when they could have theoretically shown either the same slowdown or lack thereof in that domain now that multiple generations are available for easy testing.

talldayo 2y ago
That's neither a new concept, nor unique to Apple. GCC and LLVM are both regularly updated by AMD and Intel who also optimize and patch their own microcode as well. When you compile the Linux kernel, and moreover Linux software, you will be doing it with the optimizations of your chosen architecture. For x86 that means OEMs provide optimizations for stuff like AVX and SSE, whereas Apple works in their equivalent proprietary frameworks like Accelerate. You can credit Apple (or more accurately, Lattner) with starting the trend, but Linux can be compiled using the same LLVM backend MacOS uses with whatever optimizations your OEM supports. > I remember a tweet from a couple years ago where the hardware-accelerated IPC calls for a system-wide performance gain Most kernels aren't bottlenecked by IPC throughput like XNU is. A modern Linux system might need IPC for dbus, a few journaling systems and systemd. Apple's entire kernel relies on the throughput of a message queue, so feasibly speeding that up matters quite a lot.