corelens
corelens is a Linux TUI for answering where local AI workloads are actually running:
- which physical cores and logical threads are busy;
- whether detectable Intel hybrid work is on P- or E-cores;
- the last logical CPU observed for a process or thread;
- whether Ollama models are CPU-bound, GPU-bound, or offloaded;
- all DRM GPU devices, with NVIDIA NVML metrics when available and Intel/AMD sysfs metrics where exposed.
It is deliberately not a replacement for htop or btop.
Install from crates.io
On a Linux system with Rust installed:
The binary is named corelens, so no separate launcher or configuration is
required.
┌──────────────── CPU TOPOLOGY · topology ────────────────┐
│ P0 / pkg 0 T0 82% kernel T1 31% kernel 4.2 GHz │
│ E4 / pkg 0 T4 55% capacity heuristic 3.0 GHz│
└──────────────────────────────────────────────────────────┘
┌──────────────── PROCESSES ────────────────┐ ┌── OLLAMA ──┐
│ PID NAME CPU RSS LAST CPU │ │ model │
│ 8421 *ollama 188% 1.2G 2 │ │ 100% GPU │
└───────────────────────────────────────────┘ └────────────┘
Build and run
# equivalent after installation:
Required validation commands:
CachyOS / Arch Linux
Install the Rust toolchain and build tools using the distribution packages:
GPU discovery reads /sys/class/drm/card*/device and needs no root access.
NVIDIA metrics require a working driver that exposes libnvidia-ml.so; NVML
failure does not hide Intel, AMD, or metric-less NVIDIA devices. Ollama metrics
require an Ollama server listening on the configured URL.
Permissions
No root privileges are required. corelens reads /proc, CPU sysfs topology,
optional hwmon sensors, NVML, and the local Ollama HTTP API. Processes that hide
their /proc entries or exit during a scan are skipped safely.
P-core / E-core detection
Detection is deliberately conservative:
- kernel topology
core_typemetadata when available; - reliable architecture metadata when later added;
- Linux capacity grouping;
- maximum-frequency grouping as a lower-confidence heuristic;
Unknownwhen no evidence exists.
The CPU panel shows the detection source beside logical-thread utilization. A capacity or frequency result is not presented as certain hardware identity.
Last CPU versus CPU residency
Last CPU comes from the processor field in the current /proc/<pid>/stat or
/proc/<pid>/task/<tid>/stat snapshot. It answers where the kernel most recently
reported the task, not its historical scheduler residency. Historical residency
will require a future scheduler tracing backend; this MVP does not use eBPF and
does not require root.
Limitations
- Linux only.
- CPU temperature is shown only when a CPU/package hwmon label is available.
- GPU utilization/VRAM may be unavailable when the kernel driver exposes no matching sysfs metrics or NVML cannot be loaded; discovered devices remain visible.
- Per-process GPU names are not available from every NVML driver, so GPU process rows use PID and VRAM where supported.
- Process CPU percentages begin after a second sample.
- Hybrid classification remains unknown when the kernel exposes no usable metadata.
Collector failures remain visible in the footer, while the last valid snapshot is kept for collectors that previously succeeded.
Roadmap
- scheduler tracing backend for historical CPU residency;
- richer Intel/AMD topology metadata;
- configurable process filters and process detail view;
- additional GPU backends;
- optional temperature/frequency calibration and export formats.
Controls
q / Esc quit · Tab focus · j/Down, k/Up select · Enter inspect ·
r refresh · c toggle topology/logical view · s sort processes · ? help.
License
MIT. See LICENSE.