Click a stage. The paper reports >99.9% of distance arithmetic in stages 1 and 3 goes to non-neighbors.Reported
Where the waste sits
Toy grid of candidates, 3 true neighbors. Switch between fixed 8-bit and per-candidate precision.Illustrative
Dimension slices unlock low precision
Share of sub-spaces qualifying for low precision, from the ablation.Reported: 0% and 87.5%
The runtime predictor
Five inputs, SVR with RBF kernel (evaluated via lookup table), at most 1280 training samples. A linear kernel found no low-precision opportunity.Reported
Bigger gap, fewer bits
Drag to move a cluster away from the top-k cutoff. Mapping is a conceptual sketch, not the paper's fitted model.Illustrative
Accelerator structure
The prediction module reuses the main array and priority queues; there is no second array.Per episode
Bit-plane layout
Top bit of many values stored together, then the next. Fewer bits means fewer planes fetched and fewer bit-serial cycles.Illustrative data bits
>1.18× fewer memory accesses, cluster locating~3% scheduler gain, table construction~1.15× with looser accuracy target
Bandwidth ceiling
Throughput flattens near 2000 GB/s at default settings. Curve shape is a schematic.Illustrative