_2025-02-28_14：00：26_	2025-02-28 14：00：26
tech-attached-matrix-extension@lists.riscv.org _RISC-V AME 扩展的 SiFive 提案 --- tech-attached-matrix-extension@lists.riscv.org _
TAGs:处理器 risc-v ISA Matrix
Summary: This text is a discussion between team members regarding the Zvma Attached Matrix Extension (AME) proposal. They find the specification clear and elegant, offering significant matrix computation bandwidth while maintaining software-friendliness. However, they have some concerns about physical design friendliness, specifically the placement of matrix computation logic alongside RAM/flop states and the potential routing contention between the computation block and on-chip memory. They suggest making the data layout microarchitecture (uarch) defined instead of locked at the ISA level and allowing alternative punning schemes in future extensions for greater flexibility. They also recommend specifying that matrix data should be marked as "unspecified" following any matrix configuration change. The team is considering the uarch based on the proposal and looks forward to continued collaboration as Zvma progresses toward ratification.本文是团队成员之间关于 Zvma 附加矩阵扩展 （AME） 提案的讨论。他们发现规范清晰而优雅，在保持软件友好性的同时提供了大量的矩阵计算带宽。然而，他们对物理设计友好性有一些担忧，特别是矩阵计算逻辑与 RAM/flop 状态一起放置，以及计算块和片上存储器之间潜在的路由争用。他们建议在 ISA 级别定义而不是锁定数据布局微架构 （uarch），并在未来的扩展中允许使用其他双关模式以获得更大的灵活性。他们还建议指定在矩阵配置更改后应将矩阵数据标记为“未指定”。该团队正在根据该提案考虑 uarch，并期待在 Zvma 获得批准的过程中继续合作。

_2025-02-28_14：05：35_	2025-02-28 14：05：35
tech-attached-matrix-extension@lists.riscv.org _RISC-V AME 扩展的 SiFive 提案 --- tech-attached-matrix-extension@lists.riscv.org _
TAGs:处理器 risc-v ISA Matrix
Summary: The text discusses a feedback exchange between team members regarding the Zvma Attached Matrix Extension (AME) proposal. The team, T1, has reviewed the proposal and finds it clear and elegant, offering significant matrix computation bandwidth while maintaining software-friendliness. They suggest some improvements, including making the data layout microarchitecture (uarch) defined instead of locked at the ISA level, addressing consistency challenges in multi-core scenarios, and specifying that matrix data should be marked as "unspecified" following any matrix configuration change. They also discuss concerns about physical design friendliness, specifically the placement of matrix computation logic alongside RAM/flop states and the potential routing contention between the computation block and on-chip memory. They emphasize the importance of addressing these issues to ensure high-performance computation and alignment across different uarch designs.本文讨论了团队成员之间关于 Zvma 附加矩阵扩展 （AME） 提案的反馈交流。T1 团队审查了该提案，发现它清晰而优雅，在保持软件友好性的同时提供了大量的矩阵计算带宽。他们提出了一些改进建议，包括在 ISA 级别定义而不是锁定数据布局微架构 （uarch），解决多核场景中的一致性挑战，以及指定在任何矩阵配置更改后应将矩阵数据标记为 “unspecified”。他们还讨论了对物理设计友好性的担忧，特别是矩阵计算逻辑与 RAM/flop 状态一起放置，以及计算块和片上存储器之间潜在的布线争用。他们强调了解决这些问题的重要性，以确保不同 uarch 设计之间的高性能计算和对齐。