Winbond Electronics Corporation has announced that FPGA manufacturer Gowin Semiconductor has embedded a Winbond 64 Mb HyperRAM™ fast DRAM memory device in its new GoAI 2.0 machine learning platform.
GoAI 2.0 is a complete, new hardware and software system for machine learning applications. Compatible with the TensorFlow machine learning development environment, GoAI 2.0 is aimed at edge computing applications such as smart door locks, smart speakers, voice-activated devices and smart toys.
The hardware component of the GoAI 2.0 platform, the GW1NSR4, is a system-in-package (SiP) featuring a FPGA and ARM Cortex M3 microcontroller for the machine learning application supported by Winbond’s 64 Mb HyperRAM™ supplied in known good die (KGD) format.
The Winbond DRAM based on HyperRAM™ technology is ideal for Gowin’s target applications, in which the electronics circuit needs to be made as small as possible, while providing sufficient storage and data bandwidth to support compute-intensive workloads such as keyword detection or image recognition. Winbond’s 64 Mb HyperRAM™ product has just connected 11 signal pins, so its connections to the host FPGA is minimal – the entire GW1NSR4 SiP has a footprint of just 4.2- x 4.2-mm in a BGA package. The 64 Mb memory capacity provided by the Winbond device is sufficient to run both an operating system and to concurrently operate as buffer memory for a TinyML model, or as a frame buffer.
The performance specifications of the the 64 Mb HyperRAM™ include maximum data bandwidth of 500MB/s. It also offers ultra-low power consumption in operating and hybrid sleep modes.
Jason Zhu, CEO of Gowin, said: ‘The problem which Gowin Semiconductor has solved with the GW1NSR4 is to pack a high-performance and low-power edge computing engine in a tiny package. The Winbond KGD format and HyperRAM™ memory technology are ideal for this, because the die can be embedded in the same package as the FPGA, eliminating the need for DRAM as an external component.
Winbond’s HyperRAM™ products are available for high-volume production in densities of 512 Mb, 256 Mb, 128 Mb, 64 Mb and 32 Mb.