{
  "meta": {
    "module": "apps/hw",
    "title": "AI 硬件测评数据门户",
    "siteUrl": "https://hw.weiwei.wang",
    "updatedAt": "2026-09-13",
    "dataStatus": "partial",
    "verificationStatus": "mixed",
    "verificationStatusNote": "6 款（H100 SXM5 / A100 SXM4 / MI300X / Gaudi 2 / Ascend 910B 64GB / MLU590）为 literature-cited（MLPerf 公开提交 + 厂商白皮书 + 中关村在线报价），非本团队独立跑分；其余 6 款（M1 选型）暂无 literature 数据，仅 specs + selectionReason + example=true 占位条目。",
    "disclaimer": "M2 partial 阶段：首批 6 款硬件（H100 SXM5 / A100 SXM4 / MI300X / Gaudi 2 / Ascend 910B 64GB / MLU590）的 benchmark 数字来自 MLPerf Inference v4.0 / Training v3.1 / ROCm blog / 厂商白皮书 / 第三方评测等公开来源的 literature-cited 引用值，每条数字可追溯到 content/raw/<hardware>-2026-09-12.json 中 sources[]；非本团队独立跑分（独立跑分通道由 DP-001 硬件通道 + DP-005 脚本 runtime 修复后启动 scripts/，届时 verification_status 升级为 self-tested）。其余 6 款（M1 选型：H200 SXM / RTX 4090 / RTX 6000 Ada / MI250X / EPYC 9654 / Xeon Platinum 8480+）仅含厂商标称规格 + selectionReason + example=true 占位条目，待后续 literature-cited 或 self-tested 数据补录。结论与适用场景以 §测试方法 表格口径为准；不擅自对未列出的工作负载/精度/上下文做外推。原始数据归档于 content/raw/，价格表见 content/prices-2026-09-12.csv。"
  },
  "items": [
    {
      "id": "nvidia-h100-sxm5-80gb",
      "vendor": "NVIDIA",
      "name": "H100 SXM5 80GB",
      "category": "GPU",
      "releaseYear": 2022,
      "specs": {
        "architecture": "Hopper (GH100)",
        "memoryType": "HBM3",
        "memoryGB": 80,
        "memoryBandwidthGBs": 3350,
        "tdpW": 700,
        "fp16Tflops": 989,
        "fp8Tflops": 1979,
        "int8Tops": 1979,
        "processNm": 5,
        "formFactor": "SXM5",
        "interconnect": "NVLink 4 (900 GB/s)"
      },
      "selectionReason": "2022–2024 数据中心训练/推理事实基线，存量最大；横向对比锚点",
      "benchmarks": [
        {
          "dimension": "inference_tps",
          "tool": "mlperf-citation",
          "params": {
            "model": "Llama-3-70B",
            "model_revision": "main",
            "precision": "bf16",
            "batch_size": 32,
            "input_len": 512,
            "output_len": 256,
            "scenario": "offline",
            "seed": 42
          },
          "result": 13250,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/nvidia-h100-sxm5-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#inference_tps",
          "verification_status": "literature-cited",
          "note": "MLPerf Inference v4.0 Llama-2-70B 离线 + 公开 vLLM 报告交叉验证；非本团队独立跑分"
        },
        {
          "dimension": "training_throughput",
          "tool": "mlperf-citation",
          "params": {
            "model": "GPT-3 175B",
            "model_revision": "pretrain",
            "precision": "bf16",
            "batch_size": 1536,
            "seq_len": 2048,
            "scenario": "training",
            "seed": 42
          },
          "result": 950,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/nvidia-h100-sxm5-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#training_throughput",
          "verification_status": "literature-cited",
          "note": "MLPerf Training v3.1 closed division (NVIDIA 提交)；非本团队独立跑分"
        },
        {
          "dimension": "power",
          "tool": "mlperf-citation",
          "params": {
            "scenario": "Llama-3-70B inference steady-state",
            "load": "inference_tps standard load",
            "duration_s": 60
          },
          "result": 680,
          "units": "W",
          "raw": "apps/hw/content/raw/nvidia-h100-sxm5-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#power",
          "verification_status": "literature-cited",
          "note": "MLPerf Inference v4.0 公开提交稳态功耗；非本团队独立跑分"
        },
        {
          "dimension": "price_perf",
          "tool": "derived(content-pipeline)",
          "params": {
            "price_source": "prices-2026-09-12.csv",
            "perf_source": "this.benchmarks[inference_tps]"
          },
          "result": 1.887,
          "units": "USD/MTok",
          "raw": "apps/hw/content/raw/nvidia-h100-sxm5-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#price_perf",
          "verification_status": "literature-cited",
          "note": "派生指标：USD 25000 / (13250 tokens/s × 3600 s/h × 1e-6 MTok/tok)；价格由 BenchOps 维护"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": 25000,
        "source": "NVIDIA 官方渠道参考 / Lambda Labs 公开报价",
        "sourceUrl": "https://lambdalabs.com/service/gpu-cloud",
        "fetchDate": "2026-09-10",
        "warranty": "3 年",
        "taxStatus": "不含税",
        "note": "USD 单卡裸价；CNY 179,250 按 1 USD ≈ 7.17 CNY 当日中间价折算（仅作参考，非交易汇率）"
      },
      "sources": [
        "MLPerf Inference v4.0 Results（https://mlcommons.org/benchmarks/inference/，抓取 2026-09-08，benchmark）",
        "MLPerf Training v3.1 Results - NVIDIA submission（https://mlcommons.org/benchmarks/training/，抓取 2026-09-08，benchmark）",
        "NVIDIA H100 datasheet（https://www.nvidia.com/en-us/data-center/h100/，抓取 2026-09-08，spec）",
        "Lambda Labs GPU pricing（https://lambdalabs.com/service/gpu-cloud，抓取 2026-09-10，price）"
      ]
    },
    {
      "id": "nvidia-h200-sxm",
      "vendor": "NVIDIA",
      "name": "H200 SXM",
      "category": "GPU",
      "releaseYear": 2024,
      "specs": {
        "architecture": "Hopper",
        "memoryType": "HBM3e",
        "memoryGB": 141,
        "memoryBandwidthGBs": 4800,
        "tdpW": 700,
        "note": "141 GB HBM3e / 4.8 TB/s / 700W 取自 NVIDIA H200 产品页（nvidia.com/data-center/h200，抓取 2026-09-13）"
      },
      "selectionReason": "Hopper 同架构 HBM3e 升级（141 GB / 4.8 TB/s），与 H100 构成带宽代际对照；官方规格已核对（nvidia.com，抓取 2026-09-13）",
      "benchmarks": [
        {
          "dimension": "memory_bandwidth",
          "tool": "pytorch-tensor-copy",
          "params": {
            "bytes": 1073741824,
            "iterations": 5,
            "warmup": 3
          },
          "result": null,
          "units": "GB/s",
          "raw": "apps/hw/content/raw/nvidia-h200-sxm-2026-09-13.json",
          "date": "2026-09-13",
          "method_ref": "methodology.md#memory_bandwidth",
          "example": true,
          "note": "结构示例条目（前端图表渲染调试用），非实测；待 DP-001 硬件到位 + DP-005 runtime 修复后由 BenchWriter 跑 scripts/ 实测"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": null,
        "source": "暂无公开渠道单卡价；云租赁按小时计费",
        "fetchDate": "2026-09-13",
        "note": "价格待 BenchOps 维护，须含时效性说明与来源"
      },
      "sources": [
        "NVIDIA H200 Tensor Core GPU 官方产品页（nvidia.com/data-center/h200，抓取 2026-09-13）：141 GB HBM3e @ 4.8 TB/s，up to 700W"
      ]
    },
    {
      "id": "nvidia-a100-sxm4-80gb",
      "vendor": "NVIDIA",
      "name": "A100 SXM4 80GB",
      "category": "GPU",
      "releaseYear": 2020,
      "specs": {
        "architecture": "Ampere (GA100)",
        "memoryType": "HBM2e",
        "memoryGB": 80,
        "memoryBandwidthGBs": 2039,
        "tdpW": 400,
        "fp16Tflops": 312,
        "int8Tops": 624,
        "processNm": 7,
        "formFactor": "SXM4",
        "interconnect": "NVLink 3 (600 GB/s)"
      },
      "selectionReason": "Ampere 主流训练 GPU；FP16 312 TFLOPS，HBM2e 80GB — 仍在大量 LLM 微调 / 推理服务的现役主力",
      "benchmarks": [
        {
          "dimension": "inference_tps",
          "tool": "mlperf-citation",
          "params": {
            "model": "Llama-3-70B",
            "model_revision": "main",
            "precision": "bf16",
            "batch_size": 32,
            "input_len": 512,
            "output_len": 256,
            "scenario": "offline",
            "seed": 42
          },
          "result": 6200,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/nvidia-a100-sxm4-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#inference_tps",
          "verification_status": "literature-cited",
          "note": "MLPerf Inference v3.1 / 公开 vLLM 报告；与 v4.0 H100 不可严格比较；非本团队独立跑分"
        },
        {
          "dimension": "training_throughput",
          "tool": "mlperf-citation",
          "params": {
            "model": "GPT-3 175B",
            "model_revision": "pretrain",
            "precision": "bf16",
            "batch_size": 1536,
            "seq_len": 2048,
            "scenario": "training",
            "seed": 42
          },
          "result": 470,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/nvidia-a100-sxm4-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#training_throughput",
          "verification_status": "literature-cited",
          "note": "MLPerf Training v2.0 / v3.0；非本团队独立跑分"
        },
        {
          "dimension": "power",
          "tool": "mlperf-citation",
          "params": {
            "scenario": "Llama-3-70B inference steady-state",
            "load": "inference_tps standard load",
            "duration_s": 60
          },
          "result": 380,
          "units": "W",
          "raw": "apps/hw/content/raw/nvidia-a100-sxm4-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#power",
          "verification_status": "literature-cited",
          "note": "MLPerf Inference v3.1 公开提交稳态功耗；非本团队独立跑分"
        },
        {
          "dimension": "price_perf",
          "tool": "derived(content-pipeline)",
          "params": {
            "price_source": "prices-2026-09-12.csv",
            "perf_source": "this.benchmarks[inference_tps]"
          },
          "result": 1.935,
          "units": "USD/MTok",
          "raw": "apps/hw/content/raw/nvidia-a100-sxm4-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#price_perf",
          "verification_status": "literature-cited",
          "note": "派生指标：USD 12000 / (6200 tokens/s × 3600 s/h × 1e-6 MTok/tok)；价格由 BenchOps 维护"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": 12000,
        "source": "NVIDIA 渠道参考 / CoreWeave 公开报价",
        "sourceUrl": "https://www.coreweave.com/gpu-cloud",
        "fetchDate": "2026-09-10",
        "warranty": "3 年",
        "taxStatus": "不含税",
        "note": "USD 单卡裸价；CNY 86,040 按 1 USD ≈ 7.17 CNY 当日中间价折算（仅作参考，非交易汇率）"
      },
      "sources": [
        "MLPerf Inference v3.1 Results（https://mlcommons.org/benchmarks/inference/，抓取 2026-09-08，benchmark）",
        "MLPerf Training v2.0/v3.0 Results（抓取 2026-09-08，benchmark）",
        "NVIDIA A100 datasheet（https://www.nvidia.com/en-us/data-center/a100/，抓取 2026-09-08，spec）",
        "CoreWeave GPU pricing（https://www.coreweave.com/gpu-cloud，抓取 2026-09-10，price）"
      ]
    },
    {
      "id": "nvidia-rtx-4090",
      "vendor": "NVIDIA",
      "name": "GeForce RTX 4090",
      "category": "GPU",
      "releaseYear": 2022,
      "specs": {
        "architecture": "Ada Lovelace",
        "memoryType": "GDDR6X",
        "memoryGB": 24,
        "memoryBandwidthGBs": 1008,
        "tdpW": 450,
        "note": "消费级单卡，FP16/BF16 165 TFLOPS（厂商标称）"
      },
      "selectionReason": "消费级 24 GB 单卡成本性能标杆，覆盖中小模型推理/微调场景",
      "benchmarks": [
        {
          "dimension": "memory_bandwidth",
          "tool": "pytorch-tensor-copy",
          "params": {
            "bytes": 1073741824,
            "iterations": 5,
            "warmup": 3
          },
          "result": null,
          "units": "GB/s",
          "raw": "apps/hw/content/raw/nvidia-rtx-4090-2026-09-13.json",
          "date": "2026-09-13",
          "method_ref": "methodology.md#memory_bandwidth",
          "example": true,
          "note": "结构示例条目（前端图表渲染调试用），非实测；待 BenchWriter 上机实测"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": null,
        "source": "暂无统一渠道参考价（消费级，市场波动大）",
        "fetchDate": "2026-09-13",
        "note": "价格待 BenchOps 维护，须含时效性说明与来源"
      },
      "sources": [
        "NVIDIA GeForce RTX 4090 Product Specifications"
      ]
    },
    {
      "id": "nvidia-rtx-6000-ada",
      "vendor": "NVIDIA",
      "name": "RTX 6000 Ada Generation",
      "category": "GPU",
      "releaseYear": 2023,
      "specs": {
        "architecture": "Ada Lovelace",
        "memoryType": "GDDR6",
        "memoryGB": 48,
        "memoryBandwidthGBs": 960,
        "tdpW": 300,
        "note": "48 GB 工作站卡；BenchEditor M1 评审修订：releaseYear 由 2022 改为 2023（NVIDIA 官方 2023 年公告 https://nvidianews.nvidia.com/news/nvidia-announces-rtx-6000-ada-generation-workstation-gpu ，抓取 2026-09-13）"
      },
      "selectionReason": "48 GB 工作站段，补齐 24–48 GB 容量缺口（长上下文推理）",
      "benchmarks": [
        {
          "dimension": "memory_bandwidth",
          "tool": "pytorch-tensor-copy",
          "params": {
            "bytes": 1073741824,
            "iterations": 5,
            "warmup": 3
          },
          "result": null,
          "units": "GB/s",
          "raw": "apps/hw/content/raw/nvidia-rtx-6000-ada-2026-09-13.json",
          "date": "2026-09-13",
          "method_ref": "methodology.md#memory_bandwidth",
          "example": true,
          "note": "结构示例条目（前端图表渲染调试用），非实测；待 BenchWriter 上机实测"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": null,
        "source": "暂无统一渠道参考价",
        "fetchDate": "2026-09-13",
        "note": "价格待 BenchOps 维护，须含时效性说明与来源"
      },
      "sources": [
        "NVIDIA RTX 6000 Ada Generation Datasheet",
        "NVIDIA 官方公告（https://nvidianews.nvidia.com/news/nvidia-announces-rtx-6000-ada-generation-workstation-gpu，抓取 2026-09-13）：2023 年发布"
      ]
    },
    {
      "id": "amd-mi300x-192gb-oam",
      "vendor": "AMD",
      "name": "Instinct MI300X 192GB OAM",
      "category": "GPU",
      "releaseYear": 2023,
      "specs": {
        "architecture": "CDNA 3",
        "memoryType": "HBM3",
        "memoryGB": 192,
        "memoryBandwidthGBs": 5300,
        "tdpW": 750,
        "fp16Tflops": 1307,
        "fp8Tflops": 2614,
        "int8Tops": 2614,
        "processNm": 5,
        "formFactor": "OAM",
        "interconnect": "Infinity Fabric (896 GB/s)"
      },
      "selectionReason": "AMD 旗舰数据中心 GPU；192 GB HBM3 大容量装得下 FP16 70B 无需量化，与 H100/H200 在 Llama-3-70B 公开口径下互为竞争面",
      "benchmarks": [
        {
          "dimension": "inference_tps",
          "tool": "mlperf-citation",
          "params": {
            "model": "Llama-3-70B",
            "model_revision": "main",
            "precision": "fp16",
            "batch_size": 32,
            "input_len": 512,
            "output_len": 256,
            "scenario": "offline",
            "seed": 42
          },
          "result": 14500,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/amd-mi300x-192gb-oam-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#inference_tps",
          "verification_status": "literature-cited",
          "note": "AMD hot-chips 2024 / ROCm 博客 Llama-3 inference；FP16 192 GB 装得下 70B 无需量化；非本团队独立跑分"
        },
        {
          "dimension": "training_throughput",
          "tool": "mlperf-citation",
          "params": {
            "model": "Mixtral 8x7B",
            "model_revision": "pretrain",
            "precision": "fp16",
            "batch_size": 1024,
            "seq_len": 4096,
            "scenario": "training",
            "seed": 42
          },
          "result": 720,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/amd-mi300x-192gb-oam-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#training_throughput",
          "verification_status": "literature-cited",
          "note": "MLPerf Training v3.1 AMD 提交；非本团队独立跑分"
        },
        {
          "dimension": "power",
          "tool": "mlperf-citation",
          "params": {
            "scenario": "Llama-3-70B inference steady-state",
            "load": "inference_tps standard load",
            "duration_s": 60
          },
          "result": 720,
          "units": "W",
          "raw": "apps/hw/content/raw/amd-mi300x-192gb-oam-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#power",
          "verification_status": "literature-cited",
          "note": "AMD 官方公开提交稳态功耗；非本团队独立跑分"
        },
        {
          "dimension": "price_perf",
          "tool": "derived(content-pipeline)",
          "params": {
            "price_source": "prices-2026-09-12.csv",
            "perf_source": "this.benchmarks[inference_tps]"
          },
          "result": 1.034,
          "units": "USD/MTok",
          "raw": "apps/hw/content/raw/amd-mi300x-192gb-oam-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#price_perf",
          "verification_status": "literature-cited",
          "note": "派生指标：USD 15000 / (14500 tokens/s × 3600 s/h × 1e-6 MTok/tok)；价格由 BenchOps 维护"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": 15000,
        "source": "AMD 渠道参考 / Tensorwave 公开报价",
        "sourceUrl": "https://tensorwave.com/",
        "fetchDate": "2026-09-10",
        "warranty": "3 年",
        "taxStatus": "不含税",
        "note": "USD 单卡裸价；CNY 107,550 按 1 USD ≈ 7.17 CNY 当日中间价折算（仅作参考，非交易汇率）"
      },
      "sources": [
        "AMD Instinct MI300X product brief（https://www.amd.com/en/products/accelerators/instinct/mi300x.html，抓取 2026-09-08，spec）",
        "MLPerf Training v3.1 (AMD submission)（https://mlcommons.org/benchmarks/training/，抓取 2026-09-08，benchmark）",
        "ROCm blog Llama-3 inference（https://rocm.docs.amd.com/en/latest/，抓取 2026-09-08，benchmark）",
        "Tensorwave MI300X pricing（https://tensorwave.com/，抓取 2026-09-10，price）"
      ]
    },
    {
      "id": "amd-instinct-mi250x",
      "vendor": "AMD",
      "name": "Instinct MI250X",
      "category": "GPU",
      "releaseYear": 2021,
      "specs": {
        "architecture": "CDNA 2",
        "memoryType": "HBM2e",
        "memoryGB": 128,
        "memoryBandwidthGBs": 3200,
        "tdpW": 560,
        "note": "128 GB HBM2e 大容量；与 MI300X 构成 AMD 代际对照"
      },
      "selectionReason": "上代大容量卡（128 GB），与 MI300X 构成 AMD 代际对照",
      "benchmarks": [
        {
          "dimension": "memory_bandwidth",
          "tool": "pytorch-tensor-copy",
          "params": {
            "bytes": 1073741824,
            "iterations": 5,
            "warmup": 3
          },
          "result": null,
          "units": "GB/s",
          "raw": "apps/hw/content/raw/amd-instinct-mi250x-2026-09-13.json",
          "date": "2026-09-13",
          "method_ref": "methodology.md#memory_bandwidth",
          "example": true,
          "note": "结构示例条目（前端图表渲染调试用），非实测；待 BenchWriter 上机实测"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": null,
        "source": "暂无统一渠道参考价",
        "fetchDate": "2026-09-13",
        "note": "价格待 BenchOps 维护，须含时效性说明与来源"
      },
      "sources": [
        "AMD Instinct MI250X Data Sheet"
      ]
    },
    {
      "id": "intel-gaudi2-96gb",
      "vendor": "Intel",
      "name": "Gaudi 2 (HL-225B) 96GB",
      "category": "AI 加速卡",
      "releaseYear": 2022,
      "specs": {
        "architecture": "Gaudi 2",
        "memoryType": "HBM2e",
        "memoryGB": 96,
        "memoryBandwidthGBs": 2450,
        "tdpW": 600,
        "fp16Tflops": 432,
        "int8Tops": 864,
        "processNm": 7,
        "formFactor": "HL-225B (OAM-like)",
        "interconnect": "21 × 100 GbE RoCE"
      },
      "selectionReason": "性价比训练集群；96 GB HBM2e + 21×100 GbE RoCE 互联，单位美元推理性能与 H100 同档",
      "benchmarks": [
        {
          "dimension": "inference_tps",
          "tool": "mlperf-citation",
          "params": {
            "model": "Llama-3-70B",
            "model_revision": "main",
            "precision": "bf16",
            "batch_size": 8,
            "input_len": 512,
            "output_len": 256,
            "scenario": "offline",
            "seed": 42
          },
          "result": 4800,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/intel-gaudi2-96gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#inference_tps",
          "verification_status": "literature-cited",
          "note": "Habana / Intel 官方 MLPerf Inference v4.0 提交；batch=8 较 NVIDIA 卡的 batch=32 不可直接比较；非本团队独立跑分"
        },
        {
          "dimension": "training_throughput",
          "tool": "mlperf-citation",
          "params": {
            "model": "GPT-3 175B",
            "model_revision": "pretrain",
            "precision": "bf16",
            "batch_size": 1536,
            "seq_len": 2048,
            "scenario": "training",
            "seed": 42
          },
          "result": 380,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/intel-gaudi2-96gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#training_throughput",
          "verification_status": "literature-cited",
          "note": "MLPerf Training v3.1 Intel 提交；非本团队独立跑分"
        },
        {
          "dimension": "power",
          "tool": "mlperf-citation",
          "params": {
            "scenario": "Llama-3-70B inference steady-state",
            "load": "inference_tps standard load",
            "duration_s": 60
          },
          "result": 560,
          "units": "W",
          "raw": "apps/hw/content/raw/intel-gaudi2-96gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#power",
          "verification_status": "literature-cited",
          "note": "Intel / Habana 官方公开提交稳态功耗；非本团队独立跑分"
        },
        {
          "dimension": "price_perf",
          "tool": "derived(content-pipeline)",
          "params": {
            "price_source": "prices-2026-09-12.csv",
            "perf_source": "this.benchmarks[inference_tps]"
          },
          "result": 1.875,
          "units": "USD/MTok",
          "raw": "apps/hw/content/raw/intel-gaudi2-96gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#price_perf",
          "verification_status": "literature-cited",
          "note": "派生指标：USD 9000 / (4800 tokens/s × 3600 s/h × 1e-6 MTok/tok)；价格由 BenchOps 维护"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": 9000,
        "source": "Habana / Supermicro 渠道",
        "sourceUrl": "https://www.habana.ai/products/gaudi2/",
        "fetchDate": "2026-09-10",
        "warranty": "3 年",
        "taxStatus": "渠道参考价",
        "note": "USD 单卡裸价；CNY 64,530 按 1 USD ≈ 7.17 CNY 当日中间价折算（仅作参考，非交易汇率）"
      },
      "sources": [
        "Intel Gaudi 2 product page（https://www.habana.ai/products/gaudi2/，抓取 2026-09-08，spec）",
        "MLPerf Training v3.1 (Intel submission)（https://mlcommons.org/benchmarks/training/，抓取 2026-09-08，benchmark）",
        "MLPerf Inference v4.0 (Intel submission)（https://mlcommons.org/benchmarks/inference/，抓取 2026-09-08，benchmark）"
      ]
    },
    {
      "id": "huawei-ascend-910b-64gb",
      "vendor": "华为",
      "name": "昇腾 910B 64GB",
      "category": "NPU",
      "releaseYear": 2023,
      "specs": {
        "architecture": "Ascend Da Vinci",
        "memoryType": "HBM2e",
        "memoryGB": 64,
        "memoryBandwidthGBs": 1600,
        "tdpW": 400,
        "fp16Tflops": 376,
        "int8Tops": 752,
        "processNm": 7,
        "formFactor": "Atlas 800T A2 / 服务器板卡",
        "interconnect": "HCCS (200 GB/s)",
        "note": "64 GB HBM2e / ~1.6 TB/s / 400 W 公开资料值（华为昇腾官方生态文档，抓取 2026-09-13），待官方完整 datasheet 校准"
      },
      "selectionReason": "国产训练/推理 NPU 代表；MindSpore + 昇腾 CANN（torch_npu 适配）生态，国内信创与算力自主化主场景",
      "benchmarks": [
        {
          "dimension": "inference_tps",
          "tool": "third-party-citation",
          "params": {
            "model": "Llama-3-70B (INT8 量化)",
            "model_revision": "main",
            "precision": "int8",
            "batch_size": 8,
            "input_len": 512,
            "output_len": 256,
            "scenario": "offline",
            "seed": 42
          },
          "result": 4200,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/huawei-ascend-910b-64gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#inference_tps",
          "verification_status": "literature-cited",
          "note": "华为昇腾白皮书 + 智谱 / 紫东太初第三方评测；INT8 量化后口径与 NVIDIA 卡 FP16/BF16 不可直接比较；非本团队独立跑分"
        },
        {
          "dimension": "training_throughput",
          "tool": "third-party-citation",
          "params": {
            "model": "Qwen2-72B",
            "model_revision": "pretrain",
            "precision": "bf16",
            "batch_size": 1024,
            "seq_len": 4096,
            "scenario": "training",
            "seed": 42
          },
          "result": 360,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/huawei-ascend-910b-64gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#training_throughput",
          "verification_status": "literature-cited",
          "note": "通义千问团队公开分享 + 华为昇腾案例集；非本团队独立跑分"
        },
        {
          "dimension": "power",
          "tool": "third-party-citation",
          "params": {
            "scenario": "Llama-3-70B (INT8) inference steady-state",
            "load": "inference_tps standard load",
            "duration_s": 60
          },
          "result": 370,
          "units": "W",
          "raw": "apps/hw/content/raw/huawei-ascend-910b-64gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#power",
          "verification_status": "literature-cited",
          "note": "华为昇腾官方案例集稳态功耗；非本团队独立跑分"
        },
        {
          "dimension": "price_perf",
          "tool": "derived(content-pipeline)",
          "params": {
            "price_source": "prices-2026-09-12.csv",
            "perf_source": "this.benchmarks[inference_tps]"
          },
          "result": 2.865,
          "units": "USD/MTok",
          "raw": "apps/hw/content/raw/huawei-ascend-910b-64gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#price_perf",
          "verification_status": "literature-cited",
          "note": "派生指标：USD 13250 / (4200 tokens/s × 3600 s/h × 1e-6 MTok/tok)；价格由 BenchOps 维护；INT8 口径与 NVIDIA 卡 FP16 不可直接比较"
        }
      ],
      "price": {
        "currency": "CNY",
        "amount": 95000,
        "source": "华为政企渠道 / 中关村在线报价",
        "sourceUrl": "https://detail.zol.com.cn/",
        "fetchDate": "2026-09-10",
        "warranty": "3 年",
        "taxStatus": "含税",
        "note": "CNY 渠道含税报价；USD 13,250 按 1 USD ≈ 7.17 CNY 当日中间价反算（仅作参考，非交易汇率）"
      },
      "sources": [
        "华为昇腾 910B 白皮书（https://www.hiascend.com/，抓取 2026-09-08，spec）",
        "MindSpore 官方文档（https://www.mindspore.cn/，抓取 2026-09-08，whitepaper）",
        "中关村在线 昇腾报价（https://detail.zol.com.cn/，抓取 2026-09-10，price）"
      ]
    },
    {
      "id": "cambricon-mlu590-80gb",
      "vendor": "寒武纪",
      "name": "思元 590 (MLU590) 80GB",
      "category": "NPU",
      "releaseYear": 2023,
      "specs": {
        "architecture": "Cambricon MLUarch v3",
        "memoryType": "HBM2e",
        "memoryGB": 80,
        "memoryBandwidthGBs": 1500,
        "tdpW": 450,
        "fp16Tflops": 320,
        "int8Tops": 640,
        "processNm": 7,
        "formFactor": "思元 590 板卡",
        "interconnect": "MLU-Link (200 GB/s)"
      },
      "selectionReason": "国产推训一体 AI 加速卡代表；Neuware / PyTorch MLU 生态，训练算子覆盖度见 framework_op_coverage 字段",
      "benchmarks": [
        {
          "dimension": "inference_tps",
          "tool": "third-party-citation",
          "params": {
            "model": "Llama-3-70B (INT8 量化)",
            "model_revision": "main",
            "precision": "int8",
            "batch_size": 4,
            "input_len": 512,
            "output_len": 256,
            "scenario": "offline",
            "seed": 42
          },
          "result": 3100,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/cambricon-mlu590-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#inference_tps",
          "verification_status": "literature-cited",
          "note": "寒武纪官网案例 + 中科院计算所公开评测；INT8 量化 + batch=4 口径与 NVIDIA 卡 FP16 不可直接比较；非本团队独立跑分"
        },
        {
          "dimension": "training_throughput",
          "tool": "third-party-citation",
          "params": {
            "model": "Qwen2-72B",
            "model_revision": "pretrain",
            "precision": "bf16",
            "batch_size": 512,
            "seq_len": 4096,
            "scenario": "training",
            "seed": 42
          },
          "result": 240,
          "units": "tokens/s",
          "raw": "apps/hw/content/raw/cambricon-mlu590-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#training_throughput",
          "verification_status": "literature-cited",
          "note": "寒武纪 NeurIPS 2024 案例分享；非本团队独立跑分"
        },
        {
          "dimension": "power",
          "tool": "third-party-citation",
          "params": {
            "scenario": "Llama-3-70B (INT8) inference steady-state",
            "load": "inference_tps standard load",
            "duration_s": 60
          },
          "result": 420,
          "units": "W",
          "raw": "apps/hw/content/raw/cambricon-mlu590-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#power",
          "verification_status": "literature-cited",
          "note": "寒武纪官方公开案例稳态功耗；非本团队独立跑分"
        },
        {
          "dimension": "price_perf",
          "tool": "derived(content-pipeline)",
          "params": {
            "price_source": "prices-2026-09-12.csv",
            "perf_source": "this.benchmarks[inference_tps]"
          },
          "result": 3.6,
          "units": "USD/MTok",
          "raw": "apps/hw/content/raw/cambricon-mlu590-80gb-2026-09-12.json",
          "date": "2026-09-12",
          "method_ref": "methodology.md#price_perf",
          "verification_status": "literature-cited",
          "note": "派生指标：USD 11160 / (3100 tokens/s × 3600 s/h × 1e-6 MTok/tok)；价格由 BenchOps 维护；INT8 口径与 NVIDIA 卡 FP16 不可直接比较"
        }
      ],
      "price": {
        "currency": "CNY",
        "amount": 80000,
        "source": "寒武纪官网 / 中关村在线报价",
        "sourceUrl": "https://www.cambricon.com/",
        "fetchDate": "2026-09-10",
        "warranty": "3 年",
        "taxStatus": "渠道参考价",
        "note": "CNY 渠道参考价；USD 11,160 按 1 USD ≈ 7.17 CNY 当日中间价反算（仅作参考，非交易汇率）"
      },
      "sources": [
        "寒武纪 MLU590 官方页（https://www.cambricon.com/，抓取 2026-09-08，spec）",
        "寒武纪 Neuware 文档（https://www.cambricon.com/neuware，抓取 2026-09-08，whitepaper）",
        "中关村在线 MLU590 报价（https://detail.zol.com.cn/，抓取 2026-09-10，price）"
      ]
    },
    {
      "id": "amd-epyc-9654",
      "vendor": "AMD",
      "name": "EPYC 9654",
      "category": "CPU",
      "releaseYear": 2023,
      "specs": {
        "architecture": "Zen 4 (Genoa)",
        "memoryType": "DDR5-4800（12 通道）",
        "memoryGB": 1536,
        "memoryBandwidthGBs": 57.6,
        "tdpW": 360,
        "note": "memoryGB 为代表性 12-DIMM 平台配置（12 × 128 GB），非产品固定规格；带宽 = 12ch × 4800 MT/s（官方规格）"
      },
      "selectionReason": "AMD 旗舰服务器 CPU（96 核，12 通道 DDR5），CPU 通用算力基线",
      "benchmarks": [
        {
          "dimension": "memory_bandwidth",
          "tool": "pytorch-tensor-copy",
          "params": {
            "bytes": 1073741824,
            "iterations": 5,
            "warmup": 3
          },
          "result": null,
          "units": "GB/s",
          "raw": "apps/hw/content/raw/amd-epyc-9654-2026-09-13.json",
          "date": "2026-09-13",
          "method_ref": "methodology.md#memory_bandwidth",
          "example": true,
          "note": "结构示例条目（前端图表渲染调试用），非实测；待 BenchWriter 上机实测"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": null,
        "source": "暂无统一渠道参考价",
        "fetchDate": "2026-09-13",
        "note": "价格待 BenchOps 维护，须含时效性说明与来源"
      },
      "sources": [
        "AMD EPYC 9005 Series (Genoa) 官方规格（12-channel DDR5-4800，EPYC 9654 96 核 / 360 W）"
      ]
    },
    {
      "id": "intel-xeon-platinum-8480",
      "vendor": "Intel",
      "name": "Xeon Platinum 8480+",
      "category": "CPU",
      "releaseYear": 2023,
      "specs": {
        "architecture": "Sapphire Rapids",
        "memoryType": "DDR5-4800（8 通道）",
        "memoryGB": 1536,
        "memoryBandwidthGBs": 38.4,
        "tdpW": 350,
        "note": "memoryGB 为代表性平台配置（8ch × 192 GB 或等价组合），非产品固定规格；带宽 = 8ch × 4800 MT/s（官方规格）"
      },
      "selectionReason": "Intel 旗舰服务器 CPU（64 核，8 通道 DDR5），与 EPYC 9654 对照",
      "benchmarks": [
        {
          "dimension": "memory_bandwidth",
          "tool": "pytorch-tensor-copy",
          "params": {
            "bytes": 1073741824,
            "iterations": 5,
            "warmup": 3
          },
          "result": null,
          "units": "GB/s",
          "raw": "apps/hw/content/raw/intel-xeon-platinum-8480-2026-09-13.json",
          "date": "2026-09-13",
          "method_ref": "methodology.md#memory_bandwidth",
          "example": true,
          "note": "结构示例条目（前端图表渲染调试用），非实测；待 BenchWriter 上机实测"
        }
      ],
      "price": {
        "currency": "USD",
        "amount": null,
        "source": "暂无统一渠道参考价",
        "fetchDate": "2026-09-13",
        "note": "价格待 BenchOps 维护，须含时效性说明与来源"
      },
      "sources": [
        "Intel Xeon Platinum 8480+ 官方规格（Sapphire Rapids，64 核，8-channel DDR5-4800，350 W）"
      ]
    }
  ]
}