High-performance vector searches depend on vector indexes and impose requirements on instance memories. This topic provides the minimum instance specifications for different amounts of vector data with common dimensions.
Recommended instance specifications
The following table provides recommended specifications for in-memory indexes only. If you use HGraph indexes and set the precise_io_type parameter to reader_io (a hybrid memory and disk index), you can support more vector data.
The table below lists the recommended minimum instance specifications for a single-table scenario. You can test the performance and increase the instance size to meet your requirements for queries per second (QPS) and latency.
If you perform vector searches based on exact matches, you do not need to create vector indexes. Scale out your instance specifications based on the following table.
Vector dimension | Number of rows in a vector table | Recommended minimum instance specifications |
128 | Less than 0.2 billion | 32 CPU cores |
0.2 billion to 0.4 billion | 64 CPU cores | |
Over 400 million | 128 CPU cores or above | |
256 | Less than 60 million | 32 CPU cores |
60 million to 120 million | 64 CPU cores | |
Greater than 120 million | 128 CPU cores or above | |
512 | Less than 30 million | 32 CPU cores |
30 million to 64 million | 64 CPU cores | |
Greater than 64 million | 128 CPU cores or above | |
768 | Less than 24 million | 32 CPU cores |
24 million to 48 million | 64 CPU cores | |
Greater than 48 million | 128 CPU cores or above | |
1024 | Less than 16 million | 32 CPU cores |
16 million to 32 million | 64 CPU cores | |
Greater than 32 million | 128 CPU cores or above | |
1536 | Less than 10 million | 32 CPU cores |
10 million to 20 million | 64 CPU cores | |
Greater than 20 million | 128 CPU cores or above |