From 74f0391fc23ea2d9897ef22422987bc70b4b4bb3 Mon Sep 17 00:00:00 2001
From: admin <zhyaction@foxmail.com>
Date: Thu, 3 Sep 2026 08:09:34 +0000
Subject: [PATCH] Add BMM_Theory: bmm_theory/hardware/ascend950pr.py

---
 .../bmm_theory/hardware/ascend950pr.py        | 92 +++++++++++++++++++
 1 file changed, 92 insertions(+)
 create mode 100644 BMM/BMM_Theory/bmm_theory/hardware/ascend950pr.py

diff --git a/BMM/BMM_Theory/bmm_theory/hardware/ascend950pr.py b/BMM/BMM_Theory/bmm_theory/hardware/ascend950pr.py
new file mode 100644
index 0000000..5683c07
--- /dev/null
+++ b/BMM/BMM_Theory/bmm_theory/hardware/ascend950pr.py
@@ -0,0 +1,92 @@
+"""Ascend950PR (DAV_3510) 硬件规格参数.
+
+数值来源: 《BMM算子优化分析 v0.98》§二 + 昇腾950 NPU架构白皮书.
+换芯片时逻辑结构不变, 只需新增一份同结构参数表.
+
+单位约定: 算力 FLOP/s, 带宽 Byte/s, 容量 Byte, 时延 秒.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class NpuSpec:
+    name: str = "Ascend950PR"
+
+    # ---- 核数与算力 ----
+    aic_num: int = 32                      # C: AIC (Cube) 核数
+    aiv_num: int = 64                      # AIV (Vector) 核数
+    cube_peak_tflops: float = 486.0        # 全芯片 Cube BF16 标称算力 (乘加各计一次)
+    aiv_freq_ghz: float = 1.65
+    aiv_fp32_per_cycle: int = 128          # 单 AIV 每拍 fp32 求和吞吐
+
+    # ---- 片上存储 (每核) ----
+    l1_bytes: int = 512 * 1024             # L1 Buffer 512KB
+    l0a_bytes: int = 64 * 1024             # L0A 64KB
+    l0b_bytes: int = 64 * 1024             # L0B 64KB
+    l0c_bytes: int = 256 * 1024            # L0C 256KB (fp32 累加, 4B/元素)
+
+    # ---- L2 / GM ----
+    l2_bytes: int = 128 * 1024 * 1024      # L2 Cache 128MB
+    bw_l2: float = 5.2e12                  # L2 读写带宽 5.2TB/s
+    bw_gm: float = 1.6e12                  # GM(HBM) 带宽 1.6TB/s (读写共享)
+
+    # ---- 搬移效率经验约束 (950PR 实测总结) ----
+    dvalue_recommend: int = 512            # dValue 推荐值 (Byte), 更大无额外收益
+    dvalue_min: int = 128                  # dValue 下限 (Byte)
+    dvalue_hw_min: int = 256               # DMA 硬件突发下限 (Byte) —— 尾轮文档 §2.3
+    min_tile_size: int = 16 * 1024         # min_TileSize: 单块搬移最小量 16KB
+    min_datamount_per_core: int = 480 * 1024   # min_DatamountPerCore: 单核搬移总量下限 480KB
+    min_core_num_ratio: float = 0.8        # minCoreNum ≈ 0.8 * C
+
+    # ---- DMA 固定开销 ----
+    t_cmd_ns: float = 50.0                 # T_cmd: 单次 GM->L1 DMA 命令固定开销 (ns, 估计值, 需实测标定)
+
+    # ---- Cube 计算粒度 ----
+    fractal: int = 16                      # 16x16x16 基本块
+
+    # ---- 派生量 (属性) ----
+    @property
+    def q16(self) -> float:
+        """单核 Cube BF16 峰值算力 (FLOP/s)."""
+        return self.cube_peak_tflops * 1e12 / self.aic_num
+
+    @property
+    def q_aiv(self) -> float:
+        """AIV 向量求和吞吐 (元素/s), 64 核合计."""
+        return self.aiv_num * self.aiv_fp32_per_cycle * self.aiv_freq_ghz * 1e9
+
+    @property
+    def bw_pc(self) -> float:
+        """单核 GM 带宽份额 (Byte/s) = W_GM / C = 50GB/s."""
+        return self.bw_gm / self.aic_num
+
+    @property
+    def bw_l2_pc(self) -> float:
+        """单核 L2 带宽份额 (Byte/s) = W_L2 / C ≈ 162.5GB/s."""
+        return self.bw_l2 / self.aic_num
+
+    @property
+    def r16(self) -> float:
+        """16bit 位宽平衡点算存比 R_16 = Cube峰值 / (GM带宽/2B) ≈ 607.5 FLOP/元素."""
+        return self.cube_peak_tflops * 1e12 / (self.bw_gm / 2)
+
+    @property
+    def min_core_num(self) -> int:
+        """确保高带宽利用率的并行搬移核数下限."""
+        return round(self.min_core_num_ratio * self.aic_num)
+
+    @property
+    def t_cmd(self) -> float:
+        """T_cmd (秒)."""
+        return self.t_cmd_ns * 1e-9
+
+    @property
+    def l0c_elems(self) -> int:
+        """L0C 容量 (fp32 元素数) = 65536."""
+        return self.l0c_bytes // 4
+
+
+ASCEND950PR = NpuSpec()