Work

An Accuracy Aware 8-Bit Fixed-Point MLP Accelerator with Layer-Wise Scaling for Edge ASICs

Year: 2026
Type: conference-paper
Abstract: Multi-layer perceptrons (MLPs) are widely used in edge classification, but efficient hardware deployment under aggressive quantization remains challenging. This paper presents an 8-bit fixed-point MLP accelerator with a (256–120–84–10) architecture that applies layer-wise output scaling after neuron accumulation. This controls activation dynamic range before quantization and improves inference accuracy. Using fully quantized 8-bit integer (INT8) inference, the hardware achieves 94.7% and 88% accuracy on Modified National Institute of Standards and Technology (MNIST) and Fashion-MNIST datasets, compared to Floating Point (FP32) software accuracies of 97.90% and 87.69%. The design is synthesized in a 65 nm UMC Application Specific Integrated Circuit (ASIC) process and implemented on an Artix-7 Field Programmable Gate Array (FPGA). Compared to uniform scaling, the proposed design incurs only 0.27% area, 2.0% delay, and 3.2% power overhead at 330 MHz along with only a 2.94% reduction in peak throughput due to increased critical path delay, and a 7.8% LookUp Table (LUT) and 2.3% flip-flop increase on FPGA.
Language: English
Open Access status: hybrid
OpenAlex ID: W7203636232
FWCI:
Cites: 15
Cited by:
Related to: 10