silicium369-spec/silicium-hnsw

Python

1 commits

updated Oct 16, 2026

See what people are saying

Newest first

Most points

Source	Message	Score	Date
	AVX-512 HNSW in Rust: 3.5x faster than FP32 at 99.8% recall via SQ8	3	Oct 16, 2026

README

Silicium Vector Search Evaluation Suite (x86_64 AVX-512 & AVX2)

High-throughput HNSW vector search engine written in Rust with runtime hardware dispatch (AVX-512 VNNI/FMA and AVX2 FMA), cache-aligned Flat-SoA graph layouts, and quantized INT8 SQ8 graph routing with exact FP32 top-pool reranking (4.0x primary vector memory compression).

Quick Start

# 1. Verify SHA-256 digests, C-ABI bindings, and 1-vCPU head-to-head confrontation
./reproduce_all.sh

# 2. Download and evaluate full TexMex SIFT-1M (1,000,000 vectors, 128D, 10,000 queries)
./reproduce_all.sh --sift1m

# 3. Evaluate High-Dimensional embeddings (Cohere-768D & OpenAI-1536D)
./reproduce_all.sh --high-dim

To compare directly against third-party engines (usearch, hnswlib, faiss-cpu) on the same CPU core (taskset -c 0):

pip install numpy pyyaml usearch hnswlib faiss-cpu
python3 confrontation_arena.py --num-vectors 100000 --num-queries 1000 --ef 16,32,64

1. SIFT-128D Benchmark Results (AWS `c7a.2xlarge` — AMD EPYC 9R14, Zen 4 AVX-512)

Multi-Engine 1-vCPU Confrontation (SIFT-128D, `k=10`, `M=16`, `efConstruction=100`, `taskset -c 0`)

SIFT-100K 1-vCPU Index Build Time: 5.24 s (Silicium) vs 5.93 s (usearch), 7.71 s (hnswlib), 9.32 s (FAISS)

Dataset & `efSearch`	Silicium (`QPS` / `Recall@10`)	`usearch` (`QPS` / `Recall@10`)	`hnswlib` (`QPS` / `Recall@10`)	`FAISS` (`QPS` / `Recall@10`)	Speedup vs `usearch` / `hnswlib`
SIFT-10K (`ef=12`)	190,281 QPS (`0.9931`)	146,653 QPS (`0.9701`)	134,801 QPS (`0.9705`)	135,261 QPS (`0.9689`)	1.30x / 1.41x
SIFT-10K (`ef=16`)	155,991 QPS (`0.9979`)	122,131 QPS (`0.9867`)	114,179 QPS (`0.9870`)	116,216 QPS (`0.9875`)	1.28x / 1.37x
SIFT-10K (`ef=24`)	118,027 QPS (`1.0000`)	94,813 QPS (`0.9972`)	89,588 QPS (`0.9971`)	89,930 QPS (`0.9978`)	1.24x / 1.32x
SIFT-100K (`ef=16`)	100,129 QPS (`0.9223`)	49,040 QPS (`0.9002`)	35,582 QPS (`0.9069`)	33,588 QPS (`0.8885`)	2.04x / 2.81x
SIFT-100K (`ef=32`)	55,835 QPS (`0.9736`)	29,534 QPS (`0.9696`)	21,799 QPS (`0.9720`)	19,595 QPS (`0.9640`)	1.89x / 2.56x
SIFT-100K (`ef=64`)	32,187 QPS (`0.9909`)	17,466 QPS (`0.9937`)	13,356 QPS (`0.9917`)	11,508 QPS (`0.9868`)	1.84x / 2.41x
SIFT-100K (`ef=128`)	19,027 QPS (`0.9963`)	10,644 QPS (`0.9967`)	8,385 QPS (`0.9975`)	6,968 QPS (`0.9985`)	1.79x / 2.27x

Official SIFT-1M Full-Scale Evaluation (1,000,000 vectors, 128D, 10,000 queries, `k=10`)

Default balanced Profile (M=24, efC=160): 27.79 s 8-vCPU index build (35,985 vectors/s), 85.69% Recall@10 at ef=16 (294,476 QPS) up to 99.81% Recall@10 at ef=256
High-Throughput fast_stream Profile (M=16, efC=100): 14.49 s 8-vCPU index build (68,994 vectors/s), 369,836 QPS peak throughput

`ef`	`Recall@10` (`balanced` `M=24`)	1-Core / 8-Core `QPS` (`balanced`)	1-Core `P50` / `P99` (`balanced`)	`Recall@10` (`fast_stream` `M=16`)	1-Core / 8-Core `QPS` (`fast_stream`)
`16`	0.8569 (`+9.51%`)	39,540 QPS / 294,476 QPS	`24.91µs` / `40.45µs`	`0.7618`	49,261 QPS / 369,836 QPS
`24`	0.9078 (`+7.75%`)	29,404 QPS / 219,191 QPS	`33.89µs` / `50.81µs`	`0.8303`	36,902 QPS / 278,597 QPS
`32`	0.9362 (`+6.44%`)	23,549 QPS / 175,582 QPS	`42.70µs` / `60.07µs`	`0.8718`	29,787 QPS / 224,878 QPS
`48`	0.9649 (`+4.71%`)	16,873 QPS / 125,184 QPS	`60.00µs` / `79.60µs`	`0.9178`	21,640 QPS / 162,608 QPS
`64`	0.9777 (`+3.52%`)	13,260 QPS / 99,105 QPS	`77.21µs` / `99.48µs`	`0.9425`	17,101 QPS / 128,433 QPS
`128`	0.9936 (`+1.49%`)	7,316 QPS / 54,599 QPS	`140.86µs` / `174.26µs`	`0.9787`	9,374 QPS / 71,169 QPS
`256`	0.9981 (`+0.52%`)	3,925 QPS / 29,393 QPS	`262.42µs` / `322.18µs`	`0.9929`	5,048 QPS / 38,223 QPS

2. High-Dimensional Embeddings (`OpenAI-1536D` & `Cohere-768D`, 25,000 vectors, 1,000 queries)

Quantized INT8 SQ8 graph routing with exact FP32 reranking (balanced profile, M=24, efConstruction=160) reduces primary vector memory by 4.0x (36.62 MB vs 146.48 MB at 1536D; 18.31 MB vs 73.24 MB at 768D) while accelerating both index construction and query throughput on AMD EPYC 9R14 (AWS c7a.2xlarge, 8 vCPUs, Zen 4 AVX-512):

Workload & Metric (`M=24, efC=160`)	`ef`	`Recall@10`	1-Core `QPS` (`P50` / `P99`)	8-Core `QPS`	Build (`SQ8` vs `FP32`)	Query Speedup (8c)
OpenAI-1536D (`Cosine`)	`16`	0.9761	19,935 QPS (`46.8µs` / `85.3µs`)	145,480 QPS	`0.999s` vs `3.811s` (3.81x)	3.50x
OpenAI-1536D (`Cosine`)	`32`	0.9930	13,626 QPS (`70.8µs` / `119.4µs`)	101,774 QPS	`0.999s` vs `3.811s` (3.81x)	3.32x
OpenAI-1536D (`Cosine`)	`64`	1.0000	10,322 QPS (`93.1µs` / `149.6µs`)	77,854 QPS	`0.999s` vs `3.811s` (3.81x)	3.35x
Cohere-768D (`Dot`)	`16`	0.9746	42,351 QPS (`22.7µs` / `41.6µs`)	312,248 QPS	`0.585s` vs `1.681s` (2.87x)	3.20x
Cohere-768D (`Dot`)	`32`	0.9930	26,544 QPS (`35.8µs` / `63.1µs`)	194,587 QPS	`0.585s` vs `1.681s` (2.87x)	2.41x
Cohere-768D (`Dot`)	`64`	0.9980	18,729 QPS (`51.2µs` / `84.7µs`)	141,861 QPS	`0.585s` vs `1.681s` (2.87x)	2.49x

Repository Structure & Legal

Dockerfile, config.yml, module.py: Standard ann-benchmarks container and C-ABI adapter.
confrontation_arena.py, reproduce_all.sh, scripts/download_release.sh: Standalone verification and multi-engine benchmark harness.
results/: Certified AWS c7a.2xlarge (AMD EPYC 9R14) telemetry JSON artifacts (confrontation_aws_c7a_results.json, sift1m_aws_c7a_results.json, high_dim_angular_dot_product_aws_c7a_results.json).
LICENSE, PATENTS.md, NOTICE, SECURITY.md, checksums.sha256: Evaluation license terms (SCSL-1.0 / PolyForm Noncommercial 1.0.0, Patent Pending) and SHA-256 integrity manifest.

Similar repos

silicium369-spec/silicium-hash

Silicium High-Performance SIMD Hash and SwissTable Associative Engine

same org

silicium369-spec/silicium-benchmarker

Silicium High-Performance HTTP Engine (The Benchmarker Standard)

same org

intel/ScalableVectorSearch

239

EnterpriseDB/vsbt

Vector-search benchmark toolkit

unum-cloud/usearch

Fast Open-Source Search & Clustering engine × for Vectors & Arbitrary Objects × in C++, C, Python,…

4,342

dangkhoasdc/awesome-vector-database

A curated list of awesome works related to high dimensional structure/vector search & database

362

ddickmann/colsearch

High-performance late-interaction retrieval engine for on-prem AI. ColBERT/ColPali multi-vector…

vector-index-bench/vibe

Vector Index Benchmark for Embeddings (VIBE) is an extensible benchmark for approximate nearest…

silicium369-spec/silicium-hnsw

Python

1 commits

updated Oct 16, 2026

See the code

See what people are saying

Newest first

Most points

Source	Message	Score	Date
	AVX-512 HNSW in Rust: 3.5x faster than FP32 at 99.8% recall via SQ8	3	Oct 16, 2026

README

Silicium Vector Search Evaluation Suite (x86_64 AVX-512 & AVX2)

Quick Start

# 1. Verify SHA-256 digests, C-ABI bindings, and 1-vCPU head-to-head confrontation
./reproduce_all.sh

# 2. Download and evaluate full TexMex SIFT-1M (1,000,000 vectors, 128D, 10,000 queries)
./reproduce_all.sh --sift1m

# 3. Evaluate High-Dimensional embeddings (Cohere-768D & OpenAI-1536D)
./reproduce_all.sh --high-dim

To compare directly against third-party engines (usearch, hnswlib, faiss-cpu) on the same CPU core (taskset -c 0):

pip install numpy pyyaml usearch hnswlib faiss-cpu
python3 confrontation_arena.py --num-vectors 100000 --num-queries 1000 --ef 16,32,64

1. SIFT-128D Benchmark Results (AWS `c7a.2xlarge` — AMD EPYC 9R14, Zen 4 AVX-512)

Multi-Engine 1-vCPU Confrontation (SIFT-128D, `k=10`, `M=16`, `efConstruction=100`, `taskset -c 0`)

SIFT-100K 1-vCPU Index Build Time: 5.24 s (Silicium) vs 5.93 s (usearch), 7.71 s (hnswlib), 9.32 s (FAISS)

Dataset & `efSearch`	Silicium (`QPS` / `Recall@10`)	`usearch` (`QPS` / `Recall@10`)	`hnswlib` (`QPS` / `Recall@10`)	`FAISS` (`QPS` / `Recall@10`)	Speedup vs `usearch` / `hnswlib`
SIFT-10K (`ef=12`)	190,281 QPS (`0.9931`)	146,653 QPS (`0.9701`)	134,801 QPS (`0.9705`)	135,261 QPS (`0.9689`)	1.30x / 1.41x
SIFT-10K (`ef=16`)	155,991 QPS (`0.9979`)	122,131 QPS (`0.9867`)	114,179 QPS (`0.9870`)	116,216 QPS (`0.9875`)	1.28x / 1.37x
SIFT-10K (`ef=24`)	118,027 QPS (`1.0000`)	94,813 QPS (`0.9972`)	89,588 QPS (`0.9971`)	89,930 QPS (`0.9978`)	1.24x / 1.32x
SIFT-100K (`ef=16`)	100,129 QPS (`0.9223`)	49,040 QPS (`0.9002`)	35,582 QPS (`0.9069`)	33,588 QPS (`0.8885`)	2.04x / 2.81x
SIFT-100K (`ef=32`)	55,835 QPS (`0.9736`)	29,534 QPS (`0.9696`)	21,799 QPS (`0.9720`)	19,595 QPS (`0.9640`)	1.89x / 2.56x
SIFT-100K (`ef=64`)	32,187 QPS (`0.9909`)	17,466 QPS (`0.9937`)	13,356 QPS (`0.9917`)	11,508 QPS (`0.9868`)	1.84x / 2.41x
SIFT-100K (`ef=128`)	19,027 QPS (`0.9963`)	10,644 QPS (`0.9967`)	8,385 QPS (`0.9975`)	6,968 QPS (`0.9985`)	1.79x / 2.27x

Official SIFT-1M Full-Scale Evaluation (1,000,000 vectors, 128D, 10,000 queries, `k=10`)

Default balanced Profile (M=24, efC=160): 27.79 s 8-vCPU index build (35,985 vectors/s), 85.69% Recall@10 at ef=16 (294,476 QPS) up to 99.81% Recall@10 at ef=256
High-Throughput fast_stream Profile (M=16, efC=100): 14.49 s 8-vCPU index build (68,994 vectors/s), 369,836 QPS peak throughput

`ef`	`Recall@10` (`balanced` `M=24`)	1-Core / 8-Core `QPS` (`balanced`)	1-Core `P50` / `P99` (`balanced`)	`Recall@10` (`fast_stream` `M=16`)	1-Core / 8-Core `QPS` (`fast_stream`)
`16`	0.8569 (`+9.51%`)	39,540 QPS / 294,476 QPS	`24.91µs` / `40.45µs`	`0.7618`	49,261 QPS / 369,836 QPS
`24`	0.9078 (`+7.75%`)	29,404 QPS / 219,191 QPS	`33.89µs` / `50.81µs`	`0.8303`	36,902 QPS / 278,597 QPS
`32`	0.9362 (`+6.44%`)	23,549 QPS / 175,582 QPS	`42.70µs` / `60.07µs`	`0.8718`	29,787 QPS / 224,878 QPS
`48`	0.9649 (`+4.71%`)	16,873 QPS / 125,184 QPS	`60.00µs` / `79.60µs`	`0.9178`	21,640 QPS / 162,608 QPS
`64`	0.9777 (`+3.52%`)	13,260 QPS / 99,105 QPS	`77.21µs` / `99.48µs`	`0.9425`	17,101 QPS / 128,433 QPS
`128`	0.9936 (`+1.49%`)	7,316 QPS / 54,599 QPS	`140.86µs` / `174.26µs`	`0.9787`	9,374 QPS / 71,169 QPS
`256`	0.9981 (`+0.52%`)	3,925 QPS / 29,393 QPS	`262.42µs` / `322.18µs`	`0.9929`	5,048 QPS / 38,223 QPS

2. High-Dimensional Embeddings (`OpenAI-1536D` & `Cohere-768D`, 25,000 vectors, 1,000 queries)

Workload & Metric (`M=24, efC=160`)	`ef`	`Recall@10`	1-Core `QPS` (`P50` / `P99`)	8-Core `QPS`	Build (`SQ8` vs `FP32`)	Query Speedup (8c)
OpenAI-1536D (`Cosine`)	`16`	0.9761	19,935 QPS (`46.8µs` / `85.3µs`)	145,480 QPS	`0.999s` vs `3.811s` (3.81x)	3.50x
OpenAI-1536D (`Cosine`)	`32`	0.9930	13,626 QPS (`70.8µs` / `119.4µs`)	101,774 QPS	`0.999s` vs `3.811s` (3.81x)	3.32x
OpenAI-1536D (`Cosine`)	`64`	1.0000	10,322 QPS (`93.1µs` / `149.6µs`)	77,854 QPS	`0.999s` vs `3.811s` (3.81x)	3.35x
Cohere-768D (`Dot`)	`16`	0.9746	42,351 QPS (`22.7µs` / `41.6µs`)	312,248 QPS	`0.585s` vs `1.681s` (2.87x)	3.20x
Cohere-768D (`Dot`)	`32`	0.9930	26,544 QPS (`35.8µs` / `63.1µs`)	194,587 QPS	`0.585s` vs `1.681s` (2.87x)	2.41x
Cohere-768D (`Dot`)	`64`	0.9980	18,729 QPS (`51.2µs` / `84.7µs`)	141,861 QPS	`0.585s` vs `1.681s` (2.87x)	2.49x

Repository Structure & Legal

Dockerfile, config.yml, module.py: Standard ann-benchmarks container and C-ABI adapter.
confrontation_arena.py, reproduce_all.sh, scripts/download_release.sh: Standalone verification and multi-engine benchmark harness.
results/: Certified AWS c7a.2xlarge (AMD EPYC 9R14) telemetry JSON artifacts (confrontation_aws_c7a_results.json, sift1m_aws_c7a_results.json, high_dim_angular_dot_product_aws_c7a_results.json).
LICENSE, PATENTS.md, NOTICE, SECURITY.md, checksums.sha256: Evaluation license terms (SCSL-1.0 / PolyForm Noncommercial 1.0.0, Patent Pending) and SHA-256 integrity manifest.

Similar repos

silicium369-spec/silicium-hash

Silicium High-Performance SIMD Hash and SwissTable Associative Engine

same org

silicium369-spec/silicium-benchmarker

Silicium High-Performance HTTP Engine (The Benchmarker Standard)

same org

intel/ScalableVectorSearch

239

EnterpriseDB/vsbt

Vector-search benchmark toolkit

unum-cloud/usearch

Fast Open-Source Search & Clustering engine × for Vectors & Arbitrary Objects × in C++, C, Python,…

4,342

dangkhoasdc/awesome-vector-database

A curated list of awesome works related to high dimensional structure/vector search & database

362

ddickmann/colsearch

High-performance late-interaction retrieval engine for on-prem AI. ColBERT/ColPali multi-vector…

vector-index-bench/vibe

Vector Index Benchmark for Embeddings (VIBE) is an extensible benchmark for approximate nearest…

silicium369-spec/silicium-hnsw

See what people are saying

README

Silicium Vector Search Evaluation Suite (x86_64 AVX-512 & AVX2)

Quick Start

1. SIFT-128D Benchmark Results (AWS c7a.2xlarge — AMD EPYC 9R14, Zen 4 AVX-512)

Multi-Engine 1-vCPU Confrontation (SIFT-128D, k=10, M=16, efConstruction=100, taskset -c 0)

Official SIFT-1M Full-Scale Evaluation (1,000,000 vectors, 128D, 10,000 queries, k=10)

2. High-Dimensional Embeddings (OpenAI-1536D & Cohere-768D, 25,000 vectors, 1,000 queries)

Repository Structure & Legal

Similar repos

silicium369-spec/silicium-hnsw

See what people are saying

README

Silicium Vector Search Evaluation Suite (x86_64 AVX-512 & AVX2)

Quick Start

1. SIFT-128D Benchmark Results (AWS c7a.2xlarge — AMD EPYC 9R14, Zen 4 AVX-512)

Multi-Engine 1-vCPU Confrontation (SIFT-128D, k=10, M=16, efConstruction=100, taskset -c 0)

Official SIFT-1M Full-Scale Evaluation (1,000,000 vectors, 128D, 10,000 queries, k=10)

2. High-Dimensional Embeddings (OpenAI-1536D & Cohere-768D, 25,000 vectors, 1,000 queries)

Repository Structure & Legal

Similar repos

1. SIFT-128D Benchmark Results (AWS `c7a.2xlarge` — AMD EPYC 9R14, Zen 4 AVX-512)

Multi-Engine 1-vCPU Confrontation (SIFT-128D, `k=10`, `M=16`, `efConstruction=100`, `taskset -c 0`)

Official SIFT-1M Full-Scale Evaluation (1,000,000 vectors, 128D, 10,000 queries, `k=10`)

2. High-Dimensional Embeddings (`OpenAI-1536D` & `Cohere-768D`, 25,000 vectors, 1,000 queries)

1. SIFT-128D Benchmark Results (AWS `c7a.2xlarge` — AMD EPYC 9R14, Zen 4 AVX-512)

Multi-Engine 1-vCPU Confrontation (SIFT-128D, `k=10`, `M=16`, `efConstruction=100`, `taskset -c 0`)

Official SIFT-1M Full-Scale Evaluation (1,000,000 vectors, 128D, 10,000 queries, `k=10`)

2. High-Dimensional Embeddings (`OpenAI-1536D` & `Cohere-768D`, 25,000 vectors, 1,000 queries)