Research

Paper

AI LLM March 03, 2026

Ouroboros: Wafer-Scale SRAM CIM with Token-Grained Pipelining for Large Language Model Inference

Authors

Yiqi Liu, Yudong Pan, Mengdi Wang, Shixin Zhao, Haonan Zhu, Yinhe Han, Lei Zhang, Ying Wang

Abstract

Conventional LLM inference architectures suffer from high energy and latency due to frequent data movement across memory hierarchies. We propose Ouroboros, a wafer-scale SRAM-based Computing-in-Memory (CIM) architecture that executes all operations in situ, eliminating off-chip migration. To maximize its limited first-level capacity, we introduce three innovations: Token-Grained Pipelining: Replaces sequence-level pipelining to mitigate length variations, boosting utilization and reducing activation storage. Distributed Dynamic KV Cache Management: Decouples memory from compute to leverage fragmented SRAM for efficient KV storage. Communication-Aware Mapping: Optimizes core allocation for locality and fault tolerance across the wafer. Experimental results show Ouroboros achieves average gains of $4.1\times$ in throughput and $4.2\times$ in energy efficiency, peaking at $9.1\times$ and $17\times$ for the 13B model. (*Due to the notification of arXiv "The Abstract field cannot be longer than 1,920 characters", the appeared Abstract is shortened. For the full Abstract, please download the Article.)

Metadata

arXiv ID: 2603.02737
Provider: ARXIV
Primary Category: cs.AR
Published: 2026-03-03
Fetched: 2026-03-04 03:41

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.02737v1</id>\n    <title>Ouroboros: Wafer-Scale SRAM CIM with Token-Grained Pipelining for Large Language Model Inference</title>\n    <updated>2026-03-03T08:41:33Z</updated>\n    <link href='https://arxiv.org/abs/2603.02737v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.02737v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Conventional LLM inference architectures suffer from high energy and latency due to frequent data movement across memory hierarchies. We propose Ouroboros, a wafer-scale SRAM-based Computing-in-Memory (CIM) architecture that executes all operations in situ, eliminating off-chip migration. To maximize its limited first-level capacity, we introduce three innovations:\n  Token-Grained Pipelining: Replaces sequence-level pipelining to mitigate length variations, boosting utilization and reducing activation storage. Distributed Dynamic KV Cache Management: Decouples memory from compute to leverage fragmented SRAM for efficient KV storage. Communication-Aware Mapping: Optimizes core allocation for locality and fault tolerance across the wafer.\n  Experimental results show Ouroboros achieves average gains of $4.1\\times$ in throughput and $4.2\\times$ in energy efficiency, peaking at $9.1\\times$ and $17\\times$ for the 13B model.\n  (*Due to the notification of arXiv \"The Abstract field cannot be longer than 1,920 characters\", the appeared Abstract is shortened. For the full Abstract, please download the Article.)</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AR'/>\n    <published>2026-03-03T08:41:33Z</published>\n    <arxiv:comment>17 pages, 21 figures, ASPLOS 2026</arxiv:comment>\n    <arxiv:primary_category term='cs.AR'/>\n    <author>\n      <name>Yiqi Liu</name>\n    </author>\n    <author>\n      <name>Yudong Pan</name>\n    </author>\n    <author>\n      <name>Mengdi Wang</name>\n    </author>\n    <author>\n      <name>Shixin Zhao</name>\n    </author>\n    <author>\n      <name>Haonan Zhu</name>\n    </author>\n    <author>\n      <name>Yinhe Han</name>\n    </author>\n    <author>\n      <name>Lei Zhang</name>\n    </author>\n    <author>\n      <name>Ying Wang</name>\n    </author>\n  </entry>"
}