Ilmu Komputer & AI editorial
Open AccessOA2026
Aperon Technical Report: Hierarchical No-Pointer Tangent-Local Search for High-Dimensional Approximate Nearest Neighbors
HNTL: A pointerless Block-SoA framework achieving 3.61x speedup and perfect recall with C=20 candidates on anisotropic manifold data
Yong Fu· 2026· DOI 10.48550/arXiv.2606.08813
The core problem
Approximate nearest neighbor (ANN) search in high-dimensional spaces is a foundational primitive for vector memory systems, retrieval-augmented generation, and semantic search. Proximity graph methods such as HNSW dominate the field due to their strong recall-latency trade-offs, but they incur a heavy **pointer tax**: each node stores multiple pointers to neighbors, inflating memory overhead, and graph traversal induces irregular memory accesses that stall CPU pipelines and defeat hardware prefetchers. The Aperon vector memory system addresses this with **HNTL (Hierarchical No-pointer Tangent-Local)**, a vector indexing and candidate generation framework that eliminates pointers entirely. HNTL partitions the high-dimensional space into local, coherent grains, represents vectors as low-dimensional coordinates on local tangent spaces, and scans them sequentially using a pointerless **Block-SoA (Structure-of-Arrays)** layout. This digest presents the IMRAD structure of the technical report, covering the methodology, hardware profiling results, and implications for high-dimensional ANN search.
Innovation
The report evaluates HNTL on anisotropic manifold data with and . Key findings:
- **Recall**: HNTL achieves a final **Rerank Recall@10 of 1.0000** with a candidate pool size of only vectors. This means perfect recall is obtained after reranking just 20 candidates, a dramatic reduction compared to graph-based methods that typically require hundreds of candidates.
- **Speed**: Hardware profiling via Apple kperf CPU Performance Monitoring Unit (PMU) counters demonstrates a **3.61x speedup** for the NEON auto-vectorized C++ Block-SoA scan engine over standard pointer-chasing graph traversals: **4.137 ns/vector** versus **14.951 ns/vector**.
- **Microarchitectural efficiency**: The speedup is driven by a **3.59x IPC (Instructions Per Cycle)** improvement and near-zero L1/L2 data cache misses, confirming that the pointerless layout eliminates the memory stalls that plague graph traversals.
These results are summarized in the table below:
| Metric | HNTL (Block-SoA) | Pointer-Chasing Graph |
|--------|------------------|-----------------------|
| Time per vector | 4.137 ns | 14.951 ns |
| Speedup | 3.61x | 1.0x |
| IPC | 3.59x | 1.0x |
| L1/L2 cache misses |
Approximate nearest neighbor (ANN) search in high-dimensional spaces is a foundational primitive for vector memory systems, retrieval-augmented generation, and semantic search. Proximity graph methods such as HNSW dominate the field due to their strong recall-latency trade-offs, but they incur a heavy **pointer tax**: each node stores multiple pointers to neighbors, inflating memory overhead, and graph traversal induces irregular memory accesses that stall CPU pipelines and defeat hardware prefetchers. The Aperon vector memory system addresses this with **HNTL (Hierarchical No-pointer Tangent-Local)**, a vector indexing and candidate generation framework that eliminates pointers entirely. HNTL partitions the high-dimensional space into local, coherent grains, represents vectors as low-dimensional coordinates on local tangent spaces, and scans them sequentially using a pointerless **Block-SoA (Structure-of-Arrays)** layout. This digest presents the IMRAD structure of the technical report, covering the methodology, hardware profiling results, and implications for high-dimensional ANN search.
HNTL operates in three stages: (1) **Space partitioning** into local grains via clustering or recursive splitting, ensuring each grain captures a coherent region of the data manifold. (2) **Tangent-space embedding**: for each grain, a local PCA is computed, and vectors are projected onto a low-dimensional tangent space. On anisotropic manifold data with dimension and , local PCA captures **96.3% of the variance**, enabling aggressive dimensionality reduction while preserving neighborhood structure. (3) **Pointerless Block-SoA scan**: vectors within a grain are stored in a Block-SoA layout, where coordinates are grouped by dimension rather than by vector. This allows sequential, cache-friendly scanning and auto-vectorization. The candidate generation proceeds by scanning grains in a hierarchical order, computing approximate distances in tangent space, and selecting a small candidate pool for exact reranking. The entire pipeline avoids pointer chasing, relying instead on contiguous memory blocks and SIMD-friendly loops.
Why it matters
The HNTL results highlight a fundamental trade-off in ANN search: graph-based methods achieve high recall through connectivity but pay a steep price in memory irregularity and pointer overhead. HNTL sidesteps this by exploiting **local low-dimensional structure** in high-dimensional data. The 96.3% variance captured by local PCA on anisotropic manifold data suggests that real-world embeddings often lie on or near low-dimensional manifolds, making tangent-space approximations highly effective. The pointerless Block-SoA layout is key to the 3.61x speedup: by storing coordinates dimension-wise, the scan engine can use NEON SIMD instructions and sequential prefetching, achieving 3.59x IPC and near-zero cache misses. The perfect recall with indicates that the tangent-space approximation preserves enough information to rank true neighbors highly, allowing exact reranking to recover them. Limitations include dependence on the manifold assumption and the cost of building local PCA models, which may be amortized in static or slowly evolving indexes. Future work could explore adaptive grain sizes, online updates, and integration with product quantization for further compression. Overall, HNTL offers a compelling alternative to pointer-heavy proximity graphs for high-dimensional ANN, particularly in memory-constrained or latency-sensitive vector memory systems.
Who should read this
CS practitioners and researchers
Opening member content…