GSI Technology rolls out servers for facial recognition and re-identification

GSI Technology has released Leda-E and Leda-S for the processing of huge datasets and high-performance computation in areas like facial recognition, object detection, and re-identification.

These server boards leverage the compute-in-memory architecture of the Associative Processing Unit (APU) and parallel processing capabilities to tackle complex applications. The company says that this is a differentiator from traditional processors that use sequential processing.

GSI Technology’s Gemini APU, a proprietary compute-in-memory processor, is designed to perform parallel processing and replace conventional processors, further enhancing the capabilities of Leda-E and Leda-S.

The Leda-E server has processing power capable of performing 1.2 peta operations per second. This single 2U server configuration consists of eight boards.

Meanwhile, the Leda-S server is designed for applications that prioritize both computational capacity and power efficiency. This 1U server configuration features sixteen boards, each achieving 800 tera operations per second.

As per the company’s statement, the APU architecture eliminates the input/output bottleneck between the processor and memory by conducting searches and computations directly at the location of the data. This results in high performance and efficiency.

One application that benefits from this technology is facial recognition systems. GSI Technology has developed a search engine that utilizes the Gemini APU to power its facial recognition capabilities. This search engine can handle 32-bit and 64-bit floating point feature vectors with 128 features.

According to the company, the GSI solution has a power consumption rate that is 3.5 times lower than CPU-only systems. Furthermore, it supports zero-shot learning, which allows for the recognition of new image categories by inserting the target vector into the search database.

Article Topics

biometrics at the edge  |  facial recognition  |  GSI Technology  |  object detection  |  re-identification