ANN Benchmark - MyVector vs PGVector vs MariaDB

Server

Cloud : OVH
Model Name      : AMD EPYC 9254 24-Core Processor
CPU : 48
RAM : 128GB

gist-960-euclidean

Index Build (Distance : L2/Euclidean)

Algorithm	M	efconstruction	Threads	Build Time
MyVector	24	200	1	24m 37s
MyVector	24	576	1	66m 39s
MyVector	24	800	1	89m 50s
MyVector	16	1200	24	6m 55s
MyVector	32	800	24	7m 1s
MyVector	32	1200	48	9m 16s
MyVector	12	2000	48	8m 8s
PGVector
MariaDB	24	N.A	1	120m

dbpedia-openai-1000k-angular

Index Build (Distance : Cosine)

Algorithm	M	efconstruction	Threads	Build Time
MyVector	24	200	1	36m 12s
MyVector	24	400	48	13m 1s
MariaDB	24	N.A	1	98m

MariaDB does not have a equivalent efconstruction -> https://lists.mariadb.org/hyperkitty/list/discuss@lists.mariadb.org/thread/PPRJF4JAFE3RIKMEPAFY2IUJJ4RPHPAW/

ANN Search, k = 10

Algorithm	M	efconstruction	ef search	Recall	QPS
MyVector	24	400	10	0.851	775
			20	0.929	743
			40	0.970	800
			80	0.988	604
			200	0.996	338
			400	0.998	205
			800	0.999	124
Mariadb	24	N.A	10	0.992	887
			20	0.997	452
			40	0.998	260
			80	0.999	28
			200	1.000	13
			400	1.000	3
			800	1.000	3

ANN Search, k = 100

Algorithm	M	efconstruction	ef search	Recall	QPS
MyVector	24	400	10	0.987	419
			20	0.987	427
			40	0.987	420
			80	0.987	421
			200	0.999	283
			400	1.000	178
Mariadb	24	N.A	10	1.000	12