I ran the same Qwen3.8-27B file on two NVIDIA cards. On an RTX 5090 with 32 GB, all 66 layers stayed on the GPU and decode reached 81.5 tok/s. On an RTX 4070 SUPER with 12 GB, 28 layers landed on the CPU and decode fell to 5.7 tok/s. Picchio caught the 14.3× gap.
Download picchio.pyz and run it with Python 3.9+ on macOS, Linux or Windows:
python picchio.pyz MODEL
Picchio reports GPU layer placement, prefill, decode, GPU work, memory, power and energy per token. Add --share bug-report for a paste-ready Ollama or llama.cpp Issue.