Similarity Search Overview

Fast and Accurate

The GSI APU is a true compute-in-memory device that inherently excels in associative parallel search. GSI applications can utilize this capability to support fast, 100% accurate KNN searches in the thousands of parameters per chip, or to balance the workload with the host processor to provide billion scale ANN searches per chip.

The GSI APU expands like memory in a server, so additional resources just require more APU PCIe cards to be installed in the server without expensive, high-speed interfaces between them. This allows users to tune their usage for database size, number of characteristics, accuracy, and latency depending on their application budget.

Scalable and Low Power

GSI’s similarity search engine can process 32-bit and 64-bit floating-point vectors with 128 features and above (no upper limit).

A scalable solution, large datasets and multiple queries can be processed using one or multiple APU boards, with no loss of performance and with very low power consumption—approximately 3.5 times lower power consumption when compared with CPU-only systems.

Additionally, the GSI solution supports zero-shot learning, whereby a new image category can be recognized by inserting the target vector into the search database.