NPU.

Bananas beat Paris.

A reader's comment on the last embedding post sent me back to check EmbeddingGemma on two backends, llama.cpp on the GPU and FastFlowLM on the NPU, against the reference model. Both were wrong, in different ways.