English

Mapping Gemma3 onto an Edge Dataflow Architecture

Distributed, Parallel, and Cluster Computing 2026-02-25 v2

Abstract

We present the first end-to-end deployment of the Gemma3 family of large language and vision models on a tiled edge dataflow architecture (AMD Ryzen AI NPU). Our work introduces a set of hardware-aware techniques. For prefill, we introduce an efficient dequantization engine, optimize tiled matrix multiplication kernels, and propose FlowQKV, a chunked, pipelined attention mechanism. For decoding, we introduce FusedDQP, which fuses dequantization and projection into a single kernel, and FlowKV, which re-structures attention to sustain high memory bandwidth utilization. Together with a compact Q4NX 4-bit quantization format, these methods yield up to 5.2×5.2\times faster prefill and 4.8×4.8\times faster decoding versus the iGPU, and 33.5×33.5\times and 2.2×2.2\times over the CPU, respectively. Power efficiency improves by as much as 67.2×67.2\times and 222.9×222.9\times compared to the iGPU and CPU. The proposed approach demonstrates that modern NPUs can deliver practical, low-power LLM and VLM inference at the edge, and provides a generalizable blueprint for mapping transformer-based models onto tiled dataflow accelerators.

Keywords

Cite

@article{arxiv.2602.06063,
  title  = {Mapping Gemma3 onto an Edge Dataflow Architecture},
  author = {Shouyu Du and Miaoxiang Yu and Zhenyu Xu and Zhiheng Ni and Jillian Cai and Qing Yang and Tao Wei},
  journal= {arXiv preprint arXiv:2602.06063},
  year   = {2026}
}

Comments

Original Version, data shall be updated

R2 v1 2026-07-01T10:23:11.401Z