loading status…
The corpus is FineWeb-Edu passages (~110 words each) embedded with Qwen3-Embedding-0.6B. Each SSD holds its slice twice: a 128-dim int8 key per passage (the first 128 dims of the embedding, in IVF-list order) that the FPGA engines stream through a top-k unit, and a 4 KiB record with the full 1024-dim vector, URL and text. A query is embedded on a GPU box, every FPGA returns its best 1023 keys (ivf: only the nprobe lists nearest the query; exact: its whole slice), the best candidates records are read back from the SSDs and ranked by exact 1024-dim cosine, and the top rerank passages are scored by Qwen3-Reranker-8B. The highlighted sentence is the one whose own embedding is closest to the query. No host-side vector index beyond the 65k IVF centroids. faiss mode (and the compare box) runs the same keys and lists as a faiss on-disk IVF on the host's 32 cores and one gen4 NVMe, the conventional way to do this; its scores are the same integer dot products, so the two paths can be compared hit for hit.