OCTOPAI ENGINEERING ARTS · INTERNAL · THE NPU SPECIALIST
THE NPU SPECIALIST · DAYS ON ONE DEVICE
PROVE ITbusy-time counters · R² 0.991
REAL MODELS9B to 35B · Whisper · 58/60
DECISIONSd3 lite / nano / mini · fidelity check
CONTROLSCPU · iGPU · cloud read zero
The NPU did the prefill.
The proof is in the counters.
Prove the chip did anything at all
Before any claim about the NPU, the specialist had to show it was doing the work. Busy time grows in a straight line with prompt length — and every non-NPU path reads exactly zero.
N4 · MEASURED (PRELIMINARY) · ONE DEVICE
0.377 ms PER PROMPT TOKEN · R² 0.991 · 25 REQUESTS
SOURCE · NPU STUDY · METRUM AI BENCH-CLI · PRELIMINARY, ONE DEVICE
OCTOPAI ENGINEERING ARTS · INTERNAL · ONE SENTENCE
N4 · MEASURED (PRELIMINARY) · 0.377 ms/token
NPU-ONLY CONTROL 96% PEAK · 87% SUSTAINED
SOURCE · NPU STUDY · COUNTER CALIBRATION WITHIN 0.36%
OCTOPAI ENGINEERING ARTS · INTERNAL · THE EDGE OF THE HIERARCHY
48/48
MEASURED · ROUTER ON THE NPU · 48/48 ROUTED AS LABELLED
FULL SET IN 2.9 MIN · DOWN FROM 22.1
The Router ran on the laptop
Easy prompts stayed on the laptop; hard ones went to Instinct GPUs. All forty-eight test prompts were routed exactly as labelled — the edge of the hierarchy, working.
R2 · MEASURED · 204/204 ON THE HOSTED ROUTER
90/90 ROUTER REQUESTS SUCCEEDED
LOCAL REQUESTS · $0 PROVIDER USAGE
PRELIMINARY · ONE DEVICE
SOURCE · NPU STUDY · ROUTER RUN, 30 OCT · METRUM AI BENCH-CLI