首页 > AI前沿 > AMMA: A Multi-Chiplet Memory-Centric Architecture for Low-Latency 1M Context Attention Serving

AMMA: A Multi-Chiplet Memory-Centric Architecture for Low-Latency 1M Context Attention Serving

arXiv机器学习 2026-04-29 04:36 4 阅读 查看原文
思考过程:1. 分析请求: * 角色:专业的新闻排版工程师。 * 任务:将提供的纯文本新闻正文排版成语义化的HTML片段。 * 规则: 1. 只输出HTML片段本身,不要Markdown围栏、不要解释。 2. 每个自然段落用`

`包裹,长文用`

`/`

`划分小节。 3. 列表内容用`
  • `,引用用`
    `,代码用`
    `,重要句子可用``。
            4.  保留原文所有信息,不得增删改内容。
            5.  不要输出``/``/``等文档级标签。
        *   正文内容:一段关于AMMA架构的学术论文摘要/新闻稿。
            *   句子1: All current LLM serving systems place the GPU at the center...
            *   句子2: Even academic PIM/PNM proposals still treat the GPU...
            *   句子3: Yet the GPU's compute-rich architecture is fundamentally mismatched...
            *   句子4: The problem is compounded as reasoning and agentic workloads...
            *   句子5: To address these inefficiencies, we present AMMA...
            *   句子6: AMMA replaces GPU compute dies with HBM-PNM cubes...
            *   句子7: To translate this bandwidth into proportional performance gains, we introduce (i) a logic-die microarchitecture..., (ii) a two-level hybrid parallelism scheme, and (iii) a reordered collective flow...
            *   句子8: We further conduct a design-space exploration...
            *   句子9: Evaluations show that AMMA achieves 15.5X lower attention latency and 6.9X lower energy consumption compared with the NVIDIA H100.
    
    2.  规划HTML结构:
        *   由于这是一段连续的学术摘要/新闻稿,可以将其分为几个逻辑小节(使用 `

    ` 或 `

    `)。 * 小节1:背景与当前挑战(句子1-4) * 小节2:AMMA架构的提出(句子5-6) * 小节3:技术创新与评估(句子7-9,其中句子7包含三个并列的创新点,可以用 `
    • ` 或 `` 标记,但规则说“保留原文所有信息,不得增删改内容”,如果用 `
      • ` 拆成列表,算不算“增删改”?通常把 (i)... (ii)... (iii)... 转成列表是允许的,只要文字内容不变。不过为了严格遵守“不得增删改内容”,最好保持原文的连贯性,或者用 `
          ` 但保持原文字。考虑到规则3“列表内容用`
          • `”,原文的(i)(ii)(iii)显然是列表形式,可以转换为 `
              `,但要注意标点符号的处理。或者,直接用 `

              ` 包裹,用 `` 强调重要句子。 * 让我们仔细斟酌“不得增删改内容”。如果把句子拆成 `

            • `,可能会改变句子结构。为了安全起见: * 使用 `

              ` 划分小节。 * 段落用 `

              ` 包裹。 * 重要句子(如结论、提出的架构)用 ``。 * 关于 (i) (ii) (iii),可以用 `

                ` 包起来,把文本放进去,或者就保留在段落里。为了展示排版能力,把 (i) (ii) (iii) 提取为 `
                • ` 是合理的,只要文字一字不差。 * 原文:"we introduce (i) a logic-die microarchitecture that fully exploits per-cube internal bandwidth for decode attention under a minimal power and area budget, (ii) a two-level hybrid parallelism scheme, and (iii) a reordered collective flow that reduces intra-chip die-to-die communication overhead." * 如果转为列表: we introduce: * (i) a logic-die... * (ii) a two-level... * (iii) a reordered... 这算“改”吗?严格来说改变了句法结构。为了绝对安全,不拆句,只用 `` 强调,或者用 `

                  ` 包裹整个段落。 但是规则要求“列表内容用`

                  • `”。如果原文有列表特征,我就用。原文的 (i) (ii) (iii) 是内联的。我把它拆成 `
                      ` 可能会被判定为“改内容”。所以还是保持内联,或者用 `

                      ` 包裹,里面用 ``。 等等,规则说“列表内容用

                      • ”,如果原文没有明显的列表(如1. 2. 3. 换行),就不需要强制制造列表。 * 重新规划小节: * `

                        Background and Challenges

                        ` (或者中文,但原文是英文,标题也用英文比较好,或者直接用原文的词) -> 规则没限制语言,保持英文。 * `

                        All current LLM serving systems... Yet the GPU's compute-rich architecture is fundamentally mismatched with the memory-bound nature of decode-phase attention, inflating serving latency while wasting power and die area on idle compute units. The problem is compounded as reasoning and agentic workloads push context lengths toward one million tokens, making attention latency the primary user-facing bottleneck.

                        ` * `

                        AMMA Architecture Proposal

                        ` * `

                        To address these inefficiencies, we present AMMA, a multi-chiplet, memory-centric architecture for low-latency long-context attention. AMMA replaces GPU compute dies with HBM-PNM cubes, roughly doubling the available memory bandwidth to better serve memory-bound attention workloads.

                        ` * `

                        Innovations and Evaluation

                        ` * `

                        To translate this bandwidth into proportional performance gains, we introduce (i) a logic-die microarchitecture that fully exploits per-cube internal bandwidth for decode attention under a minimal power and area budget, (ii) a two-level hybrid parallelism scheme, and (iii) a reordered collective flow that reduces intra-chip die-to-die communication overhead. We further conduct a design-space exploration over per-cube compute power and intra-chip D2D link bandwidth, providing actionable guidance for hardware designers.

                        ` * `

                        Evaluations show that AMMA achieves 15.5X lower attention latency and 6.9X lower energy consumption compared with the NVIDIA H100.

                        ` * 检查是否满足所有规则: 1. 只输出HTML片段,无Markdown,无解释。 (满足) 2. 每个自然段落用`

                        `,长文用`

                        /

                        `。 (满足,用了`

                        `和`

                        `) 3. 列表用`

                        • `,引用用`
                          `,代码用`
                          `,重要句子用``。 (原文没有明显的独立列表、引用或代码块。我用了``。如果非要用`
                            `,可以把(i)(ii)(iii)拆开,但存在修改原文排版的风险。为了展示对规则3的遵守,我可以把 (i) (ii) (iii) 转换为 `
                              ` 吗? 原文: "we introduce (i) a logic-die microarchitecture that fully exploits per-cube internal bandwidth for decode attention under a minimal power and area budget, (ii) a two-level hybrid parallelism scheme, and (iii) a reordered collective flow that reduces intra-chip die-to-die communication overhead." 如果转换为:

                              To translate this bandwidth into proportional performance gains, we introduce:

                              • (i) a logic-die microarchitecture that fully exploits per-cube internal bandwidth for decode attention under a minimal power and area budget,
                              • (ii) a two-level hybrid parallelism scheme,</