Monitor VRAM usage across your LAN

I needed one number on my desktop: how much GPU memory my Mac Studio was actually using, from another room.

My M3 Ultra sits in another room. It runs LLM inference all day. My MacBook Pro M4 Max sits at my desk. I wrote recently about how the gap between needing a tool and building one has collapsed. This was another case of that.

No app showed what I needed. Activity Monitor shows memory pressure without breaking out GPU allocation. iStatMenus focuses on CPU and thermals. ioreg will give you the raw data if you SSH in and parse it yourself. Nothing ties those numbers together in a way you can glance at from another machine.

So I built a macOS desktop widget that reads it directly and serves it over your local network.

What it shows

• VRAM allocation in GB with a color-coded pressure bar
• GPU utilization
• 60-sample rolling history
• Full system memory breakdown
• Remote hosts on your LAN

The remote part is the point. The app runs a lightweight TCP server on port 9847. On the Studio I enable the server. On my MacBook it polls every 5 seconds. Local M4 Max stats up top, remote M3 Ultra in a compact row below.

Why this metric is buried

Apple Silicon’s unified memory means the GPU doesn’t have dedicated VRAM. It shares the same pool as the CPU. The number exists inside IOKit, in the IOAccelerator driver’s PerformanceStatistics as Alloc system memory. Metal knows the total via recommendedMaxWorkingSetSize. But nothing in macOS connects those two values into something glanceable.

When you’re running a 70B model and watching memory creep toward the ceiling, the difference between 75% and 85% allocated is the difference between smooth inference and swap pressure tanking your token throughput.

Open source

Built in SwiftUI and WidgetKit. MIT licensed. If you run heavy GPU workloads on Apple Silicon, this will save you time.

GitHub: github.com/alexkerber/GPU-widget