Benchmarks¶

Results of the Criterion suite under benches/, run with make bench. The suite is run twice: once with the default features (scalar distances, single-threaded training) and once with the all feature set (SIMD distances through Hsdlib and Rayon-parallel training and batch quantization).

Test machine:

Item	Value
CPU	AMD Ryzen 5 7600X, 6 cores, 12 threads
SIMD backend	AVX-512 (reported as `Auto (AVX512VPOPCNTDQ Capable)`)
Rust	1.85.0
Vq	0.3.0
Criterion settings	0.5 s warm-up, 1.5 s measurement per benchmark

The times are Criterion's point estimates. Treat them as indicative; run make bench on your own hardware for numbers that matter to you.

Distances¶

One call to Distance::compute on two vectors of the given dimension.

Benchmark	Default features	`all` features	Speedup
squared euclidean, dim 64	24.533 ns	5.6861 ns	4.3x
euclidean, dim 64	27.666 ns	7.4838 ns	3.7x
manhattan, dim 64	26.887 ns	5.7689 ns	4.7x
cosine, dim 64	73.002 ns	10.782 ns	6.8x
squared euclidean, dim 128	58.425 ns	8.5753 ns	6.8x
euclidean, dim 128	59.325 ns	10.972 ns	5.4x
manhattan, dim 128	63.248 ns	8.8861 ns	7.1x
cosine, dim 128	156.94 ns	13.21 ns	11.9x
squared euclidean, dim 768	438.52 ns	36.788 ns	11.9x
euclidean, dim 768	491.67 ns	37.152 ns	13.2x
manhattan, dim 768	427.88 ns	32.481 ns	13.2x
cosine, dim 768	1.52 µs	49.014 ns	31.0x
squared euclidean, dim 1536	859.81 ns	72.213 ns	11.9x
euclidean, dim 1536	848.22 ns	72.863 ns	11.6x
manhattan, dim 1536	871.84 ns	65.146 ns	13.4x
cosine, dim 1536	2.7356 µs	95.798 ns	28.6x

The SIMD kernels are 4x to 30x faster than the scalar fallbacks, and the gap widens with dimension. Cosine distance gains the most because the scalar path makes three passes over the data.

Training¶

Training on 2000 vectors of dimension 128.

Benchmark	Default features	`all` features	Speedup
PQ, 8 subspaces, 256 centroids, 5 iterations	98.652 ms	27.048 ms	3.6x
TSVQ, depth 8	2.0654 ms	2.081 ms	1.0x

PQ training benefits from both parallel k-means assignment and SIMD distances. TSVQ training is dominated by variance computation and sorting, which neither feature accelerates.

Quantization¶

Single-vector calls use vectors of dimension 128. Batch calls quantize 2000 such vectors with quantize_batch.

Benchmark	Default features	`all` features	Speedup
BQ, one vector	46.801 ns	45.303 ns	1.0x
SQ, one vector	393.71 ns	386.53 ns	1.0x
PQ, one vector	11.623 µs	10.848 µs	1.1x
TSVQ, one vector	1.1805 µs	345.66 ns	3.4x
BQ, batch of 2000	120.82 µs	133.1 µs	0.9x
SQ, batch of 2000	817.97 µs	315.04 µs	2.6x
PQ, batch of 2000	23.342 ms	4.2699 ms	5.5x
TSVQ, batch of 2000	2.532 ms	256.93 µs	9.9x
PQ dequantize, batch of 2000	341.32 µs	161.6 µs	2.1x

With the all features, quantize_batch runs rows in parallel. The gain is largest for PQ and TSVQ, where each vector costs microseconds. For BQ, a single vector takes under 50 ns, so the thread scheduling overhead cancels out the parallelism and the batch call is no faster than a loop.

Persistence¶

Encoding and decoding trained quantizers with to_bytes and from_bytes. The PQ model here has 8 codebooks of 256 centroids with 16 dimensions each; the TSVQ model has depth 8 over dimension 128.

Benchmark	Default features	`all` features	Speedup
PQ to_bytes	25.073 µs	37.193 µs	0.7x
PQ from_bytes	62.962 µs	71.092 µs	0.9x
TSVQ to_bytes	38.591 µs	34.994 µs	1.1x
TSVQ from_bytes	68.018 µs	62.708 µs	1.1x

Persistence does not use SIMD or threads, so the two columns differ only by measurement noise.

Reproducing¶

cargo bench                  # Default features
cargo bench --features all   # SIMD and parallel

Criterion writes detailed reports to target/criterion/. To pass Criterion options, target one bench binary directly, for example cargo bench --bench distances -- --measurement-time 5.