INTERNAL ONLY
OCTOPAI ENGINEERING ARTS · INTERNAL · THE TECHNICAL SURPRISE
THEN THE SURPRISE
The NPU serves one
request at a time.
Nobody finds that by reading a datasheet. You find it by running two things at once — a chat model and a decision model on the same chip.
N2 · INTERNAL · PRELIMINARY · ONE DEVICE
SOURCE · NPU REPORT (PRELIMINARY) · N2 CONTENTION
OCTOPAI ENGINEERING ARTS · INTERNAL · THE TECHNICAL SURPRISE
Wait, really? Same decision, three ways
N2 · INTERNAL · MEASURED (PRELIMINARY)
KEEP THE THREE RUNS SEPARATE
LOG-SCALE BARS · MANIM
SOURCE · NPU REPORT (PRELIMINARY) · N2 · ONE DEVICE, ONE RUN EACH
OCTOPAI ENGINEERING ARTS · INTERNAL · WHAT THE FINDING CHANGED
Finding, not defect.
PRELIMINARY · MEASURED ON ONE DEVICE · LABELLED THAT WAY

CHANGED THE RECOMMENDED SPLIT

CHANGED A SLIDE

FILED UPSTREAM AS AN ISSUE
N2a / N2b / N2c · MEASURED (PRELIMINARY) · ONE DEVICE
N6 · FACT · ISSUES FILED UPSTREAM · KERNELS NOW FREE FOR COMMERCIAL USE
SOURCE · NPU STUDY · N2a / N2b / N2c · N6 · PRELIMINARY, ONE DEVICE