Historical benchmark: both lanes green
An early local run: one Lemonade endpoint serving over two compute lanes at once, iGPU llama.cpp ROCm plus NPU FLM, all on Linux. A historical local run, kept for the record, not as a claim.
The setup
One Lemonade endpoint, two compute lanes underneath it: the iGPU running llama.cpp under ROCm, and the NPU running FLM. Both on the same Linux host, both answering the same endpoint.
The run
| lane | backend | result |
|---|---|---|
| iGPU | llama.cpp ROCm | green |
| NPU | FLM | green |
Both lanes reported green in the same run. No figures were logged beyond that, which is exactly the problem with this post, and exactly why it stays historical.
Kept for the record
This is a historical local run with unsourced figures, quarantined in benchmarks/latest.json _unverified. It is kept for the record, as evidence of a setup that once worked, not as a benchmark anyone should quote.
Why it stays
The run proved the split was physically possible on one machine, and the record keeps the next attempt honest: it can point at what was green, and nothing more.