Editor’s Note: This content is contributed by Tiantian Han, Sr. Software Engineer and Tianyu Zhang, Sr. Design Engineer at Xilinx
In some resource-limited, high-performance, and low-latency scenarios, we strive for lower power consumption and higher performance without losing accuracy for AI inference. Low power consumption and high accuracy are especially critical in edge applications and low-latency ADAS. While 8-bit quantization can produce high accuracy, it requires more hardware resources. Extremely low-bit quantization, such as binary or ternary, often has a large accuracy degradation. Therefore, a full-process hardware-friendly quantization solution of 4-bit activations and 4-bit weights (4A4W) is proposed as a better accuracy/resource trade-off. With INT4 optimization, Xilinx can achieve up to a 77% performance boost on real hardware in comparison with INT8 and can achieve comparable accuracy to the full-precision models.
The whitepaper describes our implementation of a low-precision accelerator for CNN 4-bit XDPU on the Zynq® UltraScale+™ MPSoC and Zynq-7000 SoC families (16nm and 28nm), which takes full advantage of its DSP capabilities by efficiently mapping convolutional computations. This solution achieves 2X solution-level performance over the XDPU. On a 2D detection task in an ADAS system, the implementation achieves an inference speed of 230fps on a Zynq UltraScale+ MPSoC ZCU102 board, which is a 1.52X performance gain over the 8-bit XDPU.
The whitepaper focuses on:
To learn more about the INT4 Optimization solution on Xilinx devices, please download and read the following white paper.
Note: We would like to thank Dong Li, Dongliang Xie, Guangdong Liu, Lu Tian, Tiantian Han, Tianyu Zhang, and Yi Shan for their input in this white paper.
Original Date: 06-30-2020