Gpuer

SwiftUI menu bar app for monitoring macOS GPU and memory stats.

Note

This app was vibe coded using Claude Opus 4.6 and GPT-5.4. I do not have deep knowledge of macOS internals or SwiftUI such that I can confidently evaluate the end result.

Features

• "Available" headline showing how much unified memory you can still use, with a plain-English estimate of headroom
• Unified memory pool visualization showing GPU-mapped memory, apps/OS, and available space as competing claims on one shared pool — no more pretending GPU has its own VRAM
• Live Apple Silicon GPU utilization from AGXAccelerator PerformanceStatistics
• Physical footprint for per-process memory (the same metric Activity Monitor uses) instead of RSS, which inflates numbers by counting shared pages multiple times
• Swap usage with contextual explanation
• Two-minute sparklines for available memory and GPU utilization
• Two-column menu bar popover: stats on the left, process list on the right

How measurement works

Gpuer uses macOS system interfaces and command-line tools rather than private frameworks.

Memory

• Total physical memory comes from sysctl hw.memsize.
• Memory breakdown comes from host_statistics64(HOST_VM_INFO64), using page counters such as active, inactive, wired, compressed, speculative, free, and purgeable.
• Available memory is computed as total - used, where used is total - free - speculative - purgeable. This avoids double-counting purgeable and inactive pages, which can otherwise inflate the number beyond physical RAM.
• Swap usage comes from sysctl vm.swapusage.
• Memory pressure is derived from /usr/bin/memory_pressure by parsing the reported system-wide free percentage.

GPU and unified memory

On Apple Silicon, there is no separate VRAM. The CPU and GPU share the same physical memory pool. Gpuer shows this as a single unified bar rather than separate sections:

• GPU model, core count, utilization, and tracked memory come from /usr/sbin/ioreg -r -c AGXAccelerator -d 2, specifically the PerformanceStatistics dictionary.
• GPU mapped (Alloc system memory from IOKit) is the total memory the GPU driver has reserved. On machines running local AI models, this can be very large (e.g. 70 GB for a large LLM) because the model weights are memory-mapped for GPU access.
• GPU active (In use system memory from IOKit) is the subset actively being read/written by the GPU right now.
• The gap between mapped and active is memory that's allocated (often wired/pinned) but idle — for example, model weights that aren't being processed this instant.
• When GPU mapped memory is large, Gpuer explains why: this memory is your RAM shared with the GPU, not separate VRAM. This is typically the reason "wired" memory appears very high on machines running local inference.

Per-process memory

• Process list comes from ps -eo pid,rss,pcpu,comm -r.
• Each process's memory is measured using physical footprint via proc_pid_rusage(RUSAGE_INFO_V4) and the ri_phys_footprint field. This is the same metric Activity Monitor shows in its "Memory" column. It avoids the problem where RSS (Resident Set Size) double-counts shared libraries and memory-mapped files across processes.
• Processes are aggregated by executable name with a count shown (e.g. node (10)).
• If proc_pid_rusage fails for a process (e.g. insufficient permissions for system processes), Gpuer falls back to RSS from ps.

Important limitations

• GPU stats are Apple-Silicon-specific. The current implementation depends on AGXAccelerator, so non-AGX Macs may show Unknown and zeroed GPU values.
• Memory pressure is derived from free percentage, which is a weaker signal than the real VM pressure/compression behavior.
• The unified pool bar shows GPU-mapped, Apps/OS, and Available as non-overlapping segments, but in reality GPU allocations overlap with the wired memory category. The bar caps GPU-mapped at total - available to ensure it never exceeds 100%.
• The ioreg parsing is text-based, so future macOS formatting changes could break some fields.
• Process aggregation by binary name can merge unrelated processes with the same executable name.

Building

git clone https://github.com/simonw/gpuer
cd gpuer
swiftc -parse-as-library -framework SwiftUI -framework AppKit -framework IOKit -o Gpuer GpuerApp.swift
# or you can use `make`
./Gpuer

Requires macOS and Xcode command line tools (xcode-select --install).