RaBitQ-SSD: Split Codes and Pipelined I/O for SSD Vector Search
Yuexuan Xu, Jianyang Gao, Michael Norris, Alibek Zhakubayev, Pankaj Singh, Junjie Qi, Matthijs Douze, Cheng Long · NTU, ETH Zurich, Meta · 2026
arXiv:2610.02652
Query path: prove a read is unnecessary before issuing it
IVF toy: probe a few lists, not all Illustrative data
Synthetic 2-D points, 8 clusters. Recall@10 is computed on this toy set, not from the paper.
n_probe: 2
Split-RaBitQ: keep a prefix of each 1-bit code in DRAM
Suffix bits, ex-code and re-rank factors live on SSD. Grid is a D=128 schematic.
Looser bound with fewer bits Illustrative numbers
Error half-width scales as 1/√D_m (rescaled by D/D_m keeps the estimate unbiased). Positions below are made up to show the pruning test.
Memory dial at 10B × 1024-d (reported)
No lookup table: position in RAM gives the page Illustrative layout
Hover a record in memory. Records are in the same order on SSD, fixed size, never crossing a 4 KB page.
Pipeline ladder (throughput relative to Naive)
Why Overlapped collapses at 64 threads
Pages issued (reported figures, different settings)
Throughput model: QPS ≈ r_max / pages per query Illustrative grid
Only r_max = 1.63 M reads/s is measured. Other columns are hypothetical devices.
Hover a cell.
Scaling 100K → 10B (reported endpoints)
“Less than a millionth” is a denominator effect
Hold the cost at 509.4 pages (4 KB each, ≈2.09 MB) and vary the corpus. Store size = N × 1024 B, matching the 10.24 TB store. Hypothetical for N below 10B.
N =
The 3,210 QPS is the SSD ceiling ÷ pages
Caveats (click)
Pick a caveat.
References
RaBitQ-SSD — Xu, Gao, Norris, Zhakubayev, Singh, Qi, Douze, Long, 2026RaBitQ — Gao, Long, 2024Extended RaBitQ — Gao et al., 2025DiskANN — Subramanya et al., 2019SPANN — Chen et al., 2021PipeANN — Guo, Lu, 2025Starling — Wang et al., 2024Product Quantization / Quicker ADC — Jégou, Douze, Schmid; André et al.ScaNN — Guo et al., 2020Filtered-DiskANN / FreshDiskANN — Gollapudi et al.; Singh et al.HELMSMAN, AlayaLaser, VeloANN, SkipDisk — as cited in the paper (search )