SiliconScope

Why we included this project

Anyone running local models on an Apple Silicon Mac will get more detail than Activity Monitor offers: Neural Engine and Media Engine utilization, plus memory bandwidth measured against the chip's ceiling. An on-demand benchmark runs a short generation and reports tokens per second and per watt for each model, which is a solid way to compare models on your own hardware or check that a config change actually helped. A workload classifier and per-core throttling readouts make it easier to see why an on-device generation is slower than expected. Version 4 also watches other machines: install a small agent on a headless Mac mini, a Linux GPU box running Ollama, or a cloud instance, and they all show up in the same encrypted, paired dashboard, with remote Macs getting the full treatment including Neural Engine tracking. A history scrubber called Replay lets you scroll back through any metric to see what the machine was doing while a long job ran.