← sources

Research paper: A High-Throughput Custom Accelerator Architecture for Transformer Models on Edge Devices

source: semantic-scholarpublished 2026-02-20 00:00ingested 2026-07-15 07:57

retained text

Query: high bandwidth memory artificial intelligence accelerator
Authors: Ramakrishna Penaganti
Citations: 0
Transformer models have become the most popular type of architecture in deep learning today. They can do the best work in natural language processing, computer vision, and multimodal tasks. But their use on edge devices is limited by high computational costs, high memory bandwidth needs, and high energy use. This paper introduces a high-throughput custom accelerator architecture specifically engineered to facilitate efficient Transformer inference in resource-constrained edge environments. The proposed architecture incorporates a systolic tensor processing array, a specialized multi-head attention engine, and a hierarchical on-chip memory subsystem to minimize data movement and optimize parallelism. Low-precision arithmetic (INT8/INT4) and quantization-aware execution make models even more energy-efficient without losing accuracy. By using a co-optimized hardware-software stack and a pipelined execution strategy, the accelerator makes big gains in latency, throughput, and power use. Tests on typical Transformer models, such as BERT-Tiny, MobileViT-Transformer, and DistilGPT, show that they can process data up to 18.4 times faster and use 7.2 times less energy than optimized CPU and commercial NPU baselines. These results show that specialized Transformer-centric accelerators could make real-time edge intelligence possible for applications like on-device NLP, smart sensing, aerial perception, and embedded vision systems.

document fetched 2026-08-25 06:44 · canonical https://semanticscholar.org/paper/49e37cf2ce0f230376bb451394bde5c3b235115c