GOWIN Semiconductor Launches GoAI 2.0 for Embedded Machine Learning Inference

GOWIN Semiconductor Corp., the world’s fastest-growing programmable logic company, releases the latest version of their GoAI machine learning platform, providing an SDK and accelerator to perform machine learning for edge inference using convolutional neural networks on GOWIN FPGAs. GOWIN GoAI 2.0 offers direct integration into the TensorFlow and TensorFlow Lite Machine Learning Platforms, optimization for targeting GOWIN’s GW1NSR4P µSoC FPGA, and an accelerator to offload compute-intensive functions from the microcontroller embedded within GOWIN FPGAs with additional 80x performance.

Machine Learning is a rapidly developing field and development is aligning on frameworks, platforms, models, and datasets for better standardization, reliability, and ease of development. TensorFlow has become one of these aligning platforms and has included support for embedded SoC’s and microcontrollers. GoAI 2.0 adds the necessary additions to easily use TensorFlow with embedded FPGAs from GOWIN.

Recommended AI News: Platform Science Selected by Werner Enterprises to Support Innovative Driver Technology Solutions

“GoAI 2.0 has several important updates for deploying machine learning models onto edge focused, embedded FPGAs,” said Grant Jennings, Director of International Marketing for GOWIN Semiconductor. “We can use GoAI 2.0 to deploy reasonably sized and standardized machine learning models such as Mobilenet onto our GW1NSR4P µSoC FPGA. The GW1NSR4P is perfect for performing TinyML inference at the edge using TensorFlow Lite for Microcontrollers because it includes a hardened ARM Cortex-M3 microcontroller for direct model porting and control of the GoAI accelerator, 4.6K look-up tables of FPGA fabric to instantiate the GoAI 2.0 accelerator and connect sensor inputs as well as an additional 8MB of SRAM for layer storage all in a low cost 6x6mm QFN48 package. Our GoAI 2.0 SDK allows our customers to go from TensorFlow to FPGA deployment quickly and easily.”

With GoAI 2.0 no FPGA RTL or microprocessor C/C++ programming is required. The C/C++ code to drive the accelerator from the ARM Cortex-M processor is generated automatically by the GoAI 2.0 SDK. The GoAI 2.0 accelerator is offered as an FPGA IP, but is also included as part of pre-generated FPGA bitstreams with various types of sensor inputs. The GoAI 2.0 accelerator architecture is designed such that only a register map needs to be updated per model layer by the processor and no RTL changes are required to deploy or change the machine learning model used.

Recommended AI News: AdminaHealth Selected as Finalist at Insuretech Connect World Tour – Hartford

While machine learning can have a significant learning curve, GoAI 2.0 offers several starting reference designs with different types of sensor inputs to get started such as cameras, microphones, and accelerometers. Reference designs include data inference of predicting a sine wave output based on an input value, audio phrase detection for inferring a phrase of “yes” or “no” from a microphone input, and person detection inferring the presence of a human in view of a camera. Additional reference designs are continuing to be developed including gesture detection which detects the shape a user draws in the air while holding the development board using an accelerometer. Developers can get started with GoAI 2.0 using one of three GoAI Embedded Development Kits. The development kits include various sizes of GOWIN FPGAs, sensors to run the reference designs, and other peripherals such as HDMI inputs and outputs for video demonstrations.

“Machine Learning edge inference has become mainstream for applications in many marketplaces, including consumer, industrial, and medical,” said Scott Casper, Americas Director of Sales, GOWIN Semiconductor. “The GoAI 2.0 platform enables embedded system engineers to add these advanced features easily into applications at cost-effective prices. We have many resources available such as reference platforms, development kits, and field support to get engineers started right away.”

Recommended AI News: 3rd Annual Penetration Risk Report Reveals Surprising Trends, Offers New Recommendations