Research

Paper

AI LLM March 06, 2026

A Persistent-State Dataflow Accelerator for Memory-Bound Linear Attention Decode on FPGA

Authors

Neelesh Gupta, Peter Wang, Rajgopal Kannan, Viktor K. Prasanna

Abstract

Gated DeltaNet (GDN) is a linear attention mechanism that replaces the growing KV cache with a fixed-size recurrent state. Hybrid LLMs like Qwen3-Next use 75% GDN layers and achieve competitive accuracy to attention-only models. However, at batch-1, GDN decode is memory-bound on GPUs since the full recurrent state must be round-tripped through HBM every token. We show that this bottleneck is architectural, not algorithmic, as all subquadratic sequence models exhibit arithmetic intensities below 1 FLOP/B at decode time, making them more memory-bound than standard Transformers. We present an FPGA accelerator that eliminates this bottleneck by holding the full 2 MB recurrent state persistently in on-chip BRAM, converting the workload from memory-bound to compute-bound. Our design fuses the GDN recurrence into a five-phase pipelined datapath that performs only one read and one write pass over each state matrix per token, exploits Grouped Value Attention for paired-head parallelism, and overlaps preparation, computation, and output storage via dataflow pipelining. We explore four design points on an AMD Alveo U55C using Vitis HLS, varying head-level parallelism from 2 to 16 value-heads per iteration. Our fastest configuration achieves 63 $μ$s per token, 4.5$\times$ faster than the GPU reference on NVIDIA H100 PCIe. Post-implementation power analysis reports 9.96 W on-chip, yielding up to 60$\times$ greater energy efficiency per token decoded.

Metadata

arXiv ID: 2603.05931
Provider: ARXIV
Primary Category: cs.AR
Published: 2026-03-06
Fetched: 2026-03-09 06:05

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.05931v1</id>\n    <title>A Persistent-State Dataflow Accelerator for Memory-Bound Linear Attention Decode on FPGA</title>\n    <updated>2026-03-06T06:03:38Z</updated>\n    <link href='https://arxiv.org/abs/2603.05931v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.05931v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Gated DeltaNet (GDN) is a linear attention mechanism that replaces the growing KV cache with a fixed-size recurrent state. Hybrid LLMs like Qwen3-Next use 75% GDN layers and achieve competitive accuracy to attention-only models. However, at batch-1, GDN decode is memory-bound on GPUs since the full recurrent state must be round-tripped through HBM every token. We show that this bottleneck is architectural, not algorithmic, as all subquadratic sequence models exhibit arithmetic intensities below 1 FLOP/B at decode time, making them more memory-bound than standard Transformers. We present an FPGA accelerator that eliminates this bottleneck by holding the full 2 MB recurrent state persistently in on-chip BRAM, converting the workload from memory-bound to compute-bound. Our design fuses the GDN recurrence into a five-phase pipelined datapath that performs only one read and one write pass over each state matrix per token, exploits Grouped Value Attention for paired-head parallelism, and overlaps preparation, computation, and output storage via dataflow pipelining. We explore four design points on an AMD Alveo U55C using Vitis HLS, varying head-level parallelism from 2 to 16 value-heads per iteration. Our fastest configuration achieves 63 $μ$s per token, 4.5$\\times$ faster than the GPU reference on NVIDIA H100 PCIe. Post-implementation power analysis reports 9.96 W on-chip, yielding up to 60$\\times$ greater energy efficiency per token decoded.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AR'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.LG'/>\n    <published>2026-03-06T06:03:38Z</published>\n    <arxiv:comment>6 pages, 6 figures</arxiv:comment>\n    <arxiv:primary_category term='cs.AR'/>\n    <author>\n      <name>Neelesh Gupta</name>\n    </author>\n    <author>\n      <name>Peter Wang</name>\n    </author>\n    <author>\n      <name>Rajgopal Kannan</name>\n    </author>\n    <author>\n      <name>Viktor K. Prasanna</name>\n    </author>\n  </entry>"
}