Benchmark Atlas

亚洲模型实验室 · arXiv 论文

Attention Hardware Roofline

复现 Step-3 Figure 5:把 attention 计算量与 KV 访存量放到硬件 computation-bandwidth roofline 上判断瓶颈匹配。

stepfunrooflinehardwareattention
Attention compute vs KV memory access复现 Step-3 Figure 5:把 attention 计算量与 KV 访存量放到硬件 computation-bandwidth roofline 上判断瓶颈匹配。 示意数据,仅复现信息结构,不代表最新榜单结果。 Attention compute vs KV memory accessStep-3 Technical Report · Figure 5 / PDF p.8DEMO DATA00.51230150300450600H800910BA800H20DSv3 · 32KQwen3 MoE · 32KStep-3 · 32KKV memory access per token (GB) →attention compute (GFLOPs) ↑
Illustrative demo data · renderer: hardwareRoofline

Information grammar

它如何组织信息

compute-memory 坐标 + 多硬件 roofline 射线 + 8K→32K 模型轨迹 + 端点直标

Visual system
VS-76 Step-3 hardware roofline
Component ID
step3-hardware-roofline

Use when

适用判断

适合硬件感知 attention 设计;必须使用相同精度、batch 和 context 定义估算 compute 与访存。

Figure-level evidence

Figure 5 · PDF p.8

DSv3、Qwen3 MoE、Step-3 的 8K-32K 轨迹与 H800/H20/A800/910B roofline。

Verified
2026-07-13
Paper source
Step-3 Technical Report

Structured data

可替换的数据模型

示意数据,仅复现信息结构,不代表最新榜单结果。

{
  "hardware": [
    {
      "label": "H800",
      "compute": 620
    },
    {
      "label": "910B",
      "compute": 520
    },
    {
      "label": "A800",
      "compute": 430
    },
    {
      "label": "H20",
      "compute": 250
    }
  ],
  "models": [
    {
      "label": "DSv3",
      "points": [
        [
          0.25,
          145
        ],
        [
          1.15,
          590
        ]
      ]
    },
    {
      "label": "Qwen3 MoE",
      "points": [
        [
          0.55,
          25
        ],
        [
          3.15,
          100
        ]
      ]
    },
    {
      "label": "Step-3",
      "points": [
        [
          0.2,
          30
        ],
        [
          1,
          130
        ]
      ]
    }
  ]
}

Related components

同一图表家族的其他语法