ZTA-Q platform enables accurate low-precision CNN inference on RISC-V
ZTA-Q: an Open-source RISC-V Platform for Accurate Quantized CNN Inference
Hardware Architecture
Summary
Running AI models on small devices needs them to use less memory and simpler math, which can make them less accurate. The authors built ZTA-Q, a free RISC-V-based system that runs popular low-precision AI models while keeping the results nearly as accurate as usual. They also designed ways to test how simpler hardware math affects accuracy. Tests running on a real FPGA chip showed the system only slightly lowered accuracy but used extra hardware resources.
What this means in practice
- •For embedded ai developers: Deploy and evaluate TensorFlow Lite INT8 CNN models on customizable RISC-V hardware while studying accuracy trade-offs from hardware approximations.
- •For hardware accelerator designers: Test the effects of reduced-precision multipliers and simplified scaling on CNN accuracy using a configurable open RISC-V platform.
Authors
Yike Li, Ajay Kumar M, Vishnu PS, Dimitrios S. Nikolopoulos, Bo Ji, Hans Vandierendonck, Deepu John
Abstract
Low-precision inference is widely adopted in edge AI to reduce computational cost and memory footprint. However, existing open-source accelerator platforms provide limited end-to-end support for CNNs following the standard TensorFlow Lite integer inference scheme. This paper presents ZTA-Q, an open-source RISC-V-based platform that enables accurate deployment of TensorFlow Lite INT8 models. In addition to extending operator support, ZTA-Q provides a configurable post-processing datapath for studying how circuit-level approximations, including reduced multiplier precision, shared shift scaling, and simplified rounding, affect model accuracy. The proposed system is implemented on a Digilent Arty A7-100T FPGA and operates at 83.3 MHz. Evaluations on representative CNN models show that with LUT, register, and DSP overheads of 26.3%, 12.6%, and 150%, respectively, ZTA-Q limits the degradation in both top-1 and top-5 accuracy to within 0.25 percentage points.