Tsing Micro’s TX510 SoC is an advanced AI edge computing engine that has been developed for applications that include 3D sensing, face recognition, object recognition and gesture recognition. Winbond’s LPDDR3 DRAM provides bandwidth of up to 1866Mb/s and operates from a dual 1.2V/1.8V supply with power-saving features such as Deep Power-Down mode and a Clock Stop capability. When the two products are used together, they deliver the speed and accuracy required for AI imaging applications.
Tsing Micro’s TX510 SoC architecture includes a 32-bit RISC processor, a reconfigurable neural network engine, a reconfigurable general computing engine, an image signal processor and a 3D sensing engine. The TX510 SoC is then paired with Winbond’s 1Gb LPDDR3 DRAM die in a single 14mm x 14mm TFBGA System-in-Package (SiP).
The SiP is then capable of delivering a compute throughput of up to 1.2 TOPS, giving the SiP the ability to perform accurate face recognition (false acceptance ratio of 1 in 10 million) in less than 100ms, and compare features with a library of 100,000 faces in less than 50ms. Peak operating power consumption is just 450mW, and the chip uses only 0.01mW in quiescent mode.
The TX510 is intended for applications that need high-speed image detection and recognition, including biometric sensing, video surveillance, smart retail operations, smart home automation and advanced industrial automation.
Wang Bo, CEO of Tsing Micro, says: “Evaluation of the Winbond LPDDR3 DRAM shows that it is highly competitive both in terms of speed and power consumption. But equally important to us when developing the TX510 was the support and expertise that Winbond provided to us in integrating the DRAM die into a SiP in such a way that it maintained high performance and high signal integrity while taking care of thermal management.”