A system monitor that actually sees your Mac’s Neural Engine

What it does SiliconScope is a native SwiftUI dashboard and menu-bar suite for macOS 14+ on Apple Silicon that tracks CPU, GPU, memory, network, disk, battery, and—crucially—the Neural Engine, Media Engine, and memory bandwidth. It reads private but sudoless frameworks like IOReport and SMC to surface per-core temperatures, per-domain power draw, and thermal-throttling detection in a GUI meant to replace iStat Menus. It also includes an on-demand LLM benchmark that reports tokens per second and per watt-hour for local models.

The interesting bit The project is essentially a reverse-engineering exercise packaged as a daily driver: the author hand-curated per-generation SMC temperature keys for M1 through M5 chips, resolves undocumented IOReport symbols at runtime via dyld, and had to drop down to AppKit NSStatusItem because SwiftUI’s MenuBarExtra beachballs when toggled dynamically. The ANE “utilization” is honest about being a power-normalized estimate, since Apple never exposes actual Neural Engine occupancy.

Key highlights

• Sudoless access to chip-level metrics (power, residency, bandwidth) through private IOReport and SMC channels.
• AI Workload view classifies bottlenecks as bandwidth-bound, compute-bound, thermal-throttled, or memory-pressured.
• Per-metric menu-bar glyphs with live dropdowns, drawn via AppKit to work around SwiftUI scene-builder limitations.
• Curated per-generation sensor catalogs (M1–M5) for real E-core, P-core, GPU, and memory temperatures.
• Signed and notarized self-distributing app with built-in Sparkle updates; no Mac App Store sandbox.

Caveats

• ANE usage is a power-normalized estimate, not true hardware occupancy, because Apple does not expose Neural Engine utilization directly.
• Relies on private APIs (IOReport, SMC, HID), so it cannot be sandboxed and will never be on the Mac App Store.
• The build uses -undefined dynamic_lookup, meaning link errors may only surface at launch rather than build time.

Verdict Anyone running local LLMs or on-device AI on an Apple Silicon Mac who wants to see what’s actually throttling should grab this. Intel Mac owners and developers allergic to private API usage should look away.