GitHub ↗
← all posts
2026-07-06 npu benchmark correction

50 TOPS on a laptop

Benchmarking the XDNA 2 NPU. These are historical NPU benchmark notes from the first month of the project, kept because they show the work. The figures themselves are quarantined as unsourced.

The marketing number

50 TOPS is what the box promised and the reason the laptop was worth a bench at all. It is a peak figure from a datasheet, not a number we produced, and the useful question was how much of it we could reach from Linux.

The first month

The first month of notes are probe scripts, launch traces and hand-typed logs. They were written before the harness existed and before the reproducibility rules, and some of them before we understood what the counters were counting.

Reading them back now: the process was real, the evidence was not. None of those numbers has a source we can defend, so none of them ships as a claim.

What the notes attempted

measurefirst month resultstatus
decode latencyrecordedquarantined
batch decoderecordedquarantined
energy per tokenrecordedquarantined

The table is deliberately empty of real numbers. The values existed once; we chose not to reproduce them without a source.

The quarantine

These historical figures are unsourced and quarantined in benchmarks/latest.json _unverified. They are kept as process notes, not as evidence. Nothing from the first month should be quoted as a benchmark.

Why we keep them

The 50 TOPS figure survives as a target, not a boast. Keeping the bad notes is cheaper than repeating the mistakes, and every later measurement starts from a point we can see.