Apple Silicon power monitor repo released
An open GitHub power monitor for Apple Silicon Macs was shared to help track energy consumption and profile workloads, offering engineers a practical way to measure device power use. That kind of tooling makes it easier to optimize ML workloads for real thermal and battery constraints on Apple hardware (x.com).
A laptop battery is a fuel tank, but most Mac developers still build software like they cannot see the fuel gauge while the engine is running. Apple ships power tools like `powermetrics`, but older command-line monitors such as asitop rely on it and need administrator access to run. (github.com) (developer.apple.com) That gap matters on Apple Silicon because one chip is doing several jobs at once. The central processing unit, the graphics processor, and the Apple Neural Engine all sit on the same package, so a machine-learning run can heat up the same laptop you are also using for video, memory, and battery. (developer.apple.com) (github.com) The new piece of news is a GitHub project called macmon. Its repository describes it as a real-time monitor for Apple Silicon Macs from the M1 through the M5, and it says it runs without `sudo`. (github.com) macmon says it reads a private macOS application programming interface instead of shelling out to `powermetrics`. The result is a live terminal view of central processing unit, graphics processor, and Apple Neural Engine power draw without asking for root privileges first. (github.com) That changes who can actually use the tool day to day. A monitor that needs elevated permissions is awkward inside automated tests, remote sessions, and shared development machines, while macmon also exposes JavaScript Object Notation output and an Hypertext Transfer Protocol metrics server for Prometheus-style scraping. (github.com) (prometheus.io) The repository also shows the kind of numbers engineers usually have to guess at. It reports average central processing unit and graphics processor temperatures, random access memory and swap usage, per-cluster central processing unit utilization, and timestamped samples that can be piped into other tools. (github.com) That is especially useful for local machine-learning workloads because “fast” and “efficient” are not the same thing on a fanless or battery-powered Mac. A model that finishes one benchmark first can still burn more watts, trigger thermal throttling, or drain a battery sooner than a slightly slower run. (github.com) (developer.apple.com) Apple already gives developers some power data after the fact through MetricKit and App Store Connect reports. Those systems are aggregated and delayed, while a terminal monitor is immediate enough to catch a bad kernel launch, an overactive polling loop, or a graphics-heavy inference path while it is happening. (developer.apple.com 1) (developer.apple.com 2) The bigger pattern is that Apple Silicon performance work is becoming less about raw benchmark scores and more about watts per task. When a public tool can show live power on the central processing unit, graphics processor, and Apple Neural Engine at one-second intervals, engineers can tune for the machine people actually carry around instead of the benchmark chart they post online. (github.com)
