{
  "$comment": "Key facts published by Xinyuanyi. Quote every figure together with its baseline and report ID. Different baselines and different platforms must not be merged.",
  "updated": "2026-11-12",
  "organization": {
    "name": {
      "zh": "芯元一科技",
      "en": "Xinyuanyi Technology"
    },
    "legalName": {
      "zh": "芯元一（天津）半导体设备有限公司",
      "en": "Xinyuanyi (Tianjin) Semiconductor Equipment Co., Ltd."
    },
    "url": "https://vuke.top",
    "location": {
      "zh": "中国 · 天津",
      "en": "Tianjin, China"
    },
    "contact": {
      "person": {
        "zh": "王利民",
        "en": "Wang Limin"
      },
      "phone": "+8613161818898",
      "email": "mingxin@agentmail.to"
    }
  },
  "metrics": [
    {
      "id": "throughput",
      "value": "+29–40%",
      "label": {
        "zh": "推理吞吐提升",
        "en": "Higher inference throughput"
      },
      "baseline": {
        "zh": "对比本地 NVMe 单盘",
        "en": "vs. a single local NVMe drive"
      },
      "detail": {
        "zh": "Qwen3-Coder-480B-FP8、TP=8 生产部署形态，长上下文冷恢复负载：并发 8 时 +29%，并发 16（实测中提升幅度最大的一档）时 +40%；TP4×2 双实例全机口径为 +35–36%。",
        "en": "Qwen3-Coder-480B-FP8 in TP=8 production topology under a long-context cold-recovery load: +29% at concurrency 8 and +40% at concurrency 16, the operating point with the largest gain; +35–36% for the whole machine in the dual-instance TP4×2 layout."
      },
      "provenance": "measured",
      "reports": [
        "R2",
        "R3"
      ]
    },
    {
      "id": "ttft",
      "value": "−26–32%",
      "label": {
        "zh": "首 token 延迟（TTFT）降低",
        "en": "Lower time to first token (TTFT)"
      },
      "baseline": {
        "zh": "对比本地 NVMe 单盘",
        "en": "vs. a single local NVMe drive"
      },
      "detail": {
        "zh": "同一 480B·TP8 负载的三档并发下，TTFT p50 从 10.17–35.73 s 降到 7.53–26.35 s。",
        "en": "Across three concurrency levels of the same 480B TP8 load, TTFT p50 falls from 10.17–35.73 s to 7.53–26.35 s."
      },
      "provenance": "measured",
      "reports": [
        "R2"
      ]
    },
    {
      "id": "recompute",
      "value": "8.6–20×",
      "label": {
        "zh": "相对无外存重算的加速",
        "en": "Faster than recomputing without external storage"
      },
      "baseline": {
        "zh": "对比无外存、整段重算",
        "en": "vs. full recompute with no external storage"
      },
      "detail": {
        "zh": "并发 16：整段重算的 TTFT p50 为 149.5 s，接入 FX100 后为 11.85 s；吞吐 4.1 tok/s 对 74.9 tok/s。",
        "en": "At concurrency 16, TTFT p50 is 149.5 s when recomputing and 11.85 s with FX100; throughput is 4.1 tok/s versus 74.9 tok/s."
      },
      "provenance": "measured",
      "reports": [
        "R2"
      ]
    },
    {
      "id": "parallel-read",
      "value": "4.1×",
      "label": {
        "zh": "LMCache 并行读补丁的 TTFT 改善",
        "en": "TTFT gain from the LMCache parallel-read patch"
      },
      "baseline": {
        "zh": "对比打补丁之前",
        "en": "vs. before the patch"
      },
      "detail": {
        "zh": "单卡、并发 16、冷读盘（Qwen2.5-32B）：TTFT 37.97 s → 9.30 s；读带宽 0.98 → 5.23 GB/s，约 5.3×。",
        "en": "One GPU, concurrency 16, cold read from disk (Qwen2.5-32B): TTFT 37.97 s → 9.30 s; read bandwidth 0.98 → 5.23 GB/s, about 5.3×."
      },
      "provenance": "measured",
      "reports": [
        "R1"
      ]
    },
    {
      "id": "model-load",
      "value": "6.2–9.3×",
      "label": {
        "zh": "模型服务加载提速",
        "en": "Faster model loading"
      },
      "baseline": {
        "zh": "对比 NFS，昇腾 910B 平台",
        "en": "vs. NFS, on Huawei Ascend 910B"
      },
      "detail": {
        "zh": "华为 Atlas 910B 平台：DeepSeek-32B 服务加载 691 s → 112 s（6.2×），DeepSeek-70B 1399 s → 150 s（9.3×）。",
        "en": "On Huawei Atlas 910B: DeepSeek-32B serving load 691 s → 112 s (6.2×); DeepSeek-70B 1399 s → 150 s (9.3×)."
      },
      "provenance": "measured",
      "reports": [
        "R9"
      ]
    },
    {
      "id": "checkpoint",
      "value": "1.9×",
      "label": {
        "zh": "训练 Checkpoint 保存提速",
        "en": "Faster training checkpoint saves"
      },
      "baseline": {
        "zh": "对比 R1 报告中的对照基线",
        "en": "vs. the baseline in report R1"
      },
      "detail": {
        "zh": "8 卡 32B LoRA，每份整模型快照 65.6 GB：178 s → 94 s；持续写带宽 3.26 → 6.40 GB/s（+96%）。",
        "en": "8-GPU 32B LoRA with 65.6 GB full-model snapshots: 178 s → 94 s; sustained write bandwidth 3.26 → 6.40 GB/s (+96%)."
      },
      "provenance": "measured",
      "reports": [
        "R1"
      ]
    },
    {
      "id": "utilization",
      "value": "46.7% → 62.8%",
      "label": {
        "zh": "模型切换场景的算力有效利用率",
        "en": "Effective compute utilization under model switching"
      },
      "baseline": {
        "zh": "每小时切换 20 次的档位",
        "en": "at 20 model switches per hour"
      },
      "detail": {
        "zh": "按实测加载时间建模：每小时切换 20 次模型时，算力有效利用率从 46.7% 提高到 62.8%。",
        "en": "Modeled from measured load times: at 20 model switches per hour, effective compute utilization rises from 46.7% to 62.8%."
      },
      "provenance": "measured",
      "reports": [
        "R1"
      ]
    },
    {
      "id": "concurrent-load",
      "value": "16% / 30%",
      "label": {
        "zh": "8 卡并发加载模型",
        "en": "Concurrent model loading on 8 GPUs"
      },
      "baseline": {
        "zh": "对比 R1 报告中的对照基线",
        "en": "vs. the baseline in report R1"
      },
      "detail": {
        "zh": "8 张卡同时冷读模型：整体墙钟时间快 16%，单个模型加载快 30%。",
        "en": "Eight GPUs cold-reading models at once: 16% faster wall-clock overall and 30% faster per model."
      },
      "provenance": "measured",
      "reports": [
        "R1"
      ]
    },
    {
      "id": "repeatability",
      "value": "≤5%",
      "label": {
        "zh": "两轮独立测量的偏差",
        "en": "Deviation between two independent runs"
      },
      "baseline": {
        "zh": "同一档位重复测量",
        "en": "same load level, measured twice"
      },
      "detail": {
        "zh": "同一档位做两轮独立测量，偏差不超过 5%；fs:// 共享池的跨实例热共享同时验证通过。",
        "en": "Two independent runs at the same load level deviate by no more than 5%; cross-instance hot sharing through the fs:// shared pool was verified in the same campaign."
      },
      "provenance": "measured",
      "reports": [
        "R3"
      ]
    },
    {
      "id": "hbm-equivalence",
      "value": "128 GB",
      "label": {
        "zh": "外置 KV 层可替代的显存驻留量级",
        "en": "HBM residency an external KV tier can stand in for"
      },
      "baseline": {
        "zh": "沐曦 N260 单卡，14B 模型七组对照",
        "en": "MetaX N260 single GPU, 14B model, seven-way comparison"
      },
      "detail": {
        "zh": "14B 模型七组对照的结论：外置 KV 层可以替代 128 GB 级别的显存驻留。该轮平台是沐曦 N260 单卡，方法论可跨平台使用，数字不与 MI308X 混用。",
        "en": "Conclusion of a seven-way comparison on a 14B model: an external KV tier can stand in for 128 GB-class HBM residency. That campaign ran on a single MetaX N260 GPU; the method carries across platforms, the numbers are not mixed with MI308X results."
      },
      "provenance": "measured",
      "reports": [
        "R5"
      ]
    },
    {
      "id": "multimodal",
      "value": "7 / 7",
      "label": {
        "zh": "ComfyUI + LTX-Video 2.3 模型实跑通过",
        "en": "ComfyUI + LTX-Video 2.3 models running end to end"
      },
      "baseline": {
        "zh": "AMD MI308X，ROCm 7.2",
        "en": "AMD MI308X, ROCm 7.2"
      },
      "detail": {
        "zh": "交付的 7 个模型（3 个 VAE、fp8 文本编码器、文本投影、2 个 LoRA）在 MI308X / ROCm 7.2 上全部实跑通过，没有出现 dtype、算子或内核错误。",
        "en": "All 7 delivered models (3 VAEs, the fp8 text encoder, the text projection and 2 LoRAs) run end to end on MI308X / ROCm 7.2 with no dtype, operator or kernel errors."
      },
      "provenance": "measured",
      "reports": [
        "R6",
        "R7"
      ]
    },
    {
      "id": "storage-ratio",
      "value": "1 : 8",
      "label": {
        "zh": "阵列与推理节点的配比",
        "en": "Arrays to inference nodes"
      },
      "baseline": {
        "zh": "生产配比口径",
        "en": "production sizing basis"
      },
      "detail": {
        "zh": "每台 FX100 服务 8 个推理节点：100 GbE 能支撑 8 节点 KV 回读是实测锚点。联测口径下，存储投入约占 GPU 预算的一成以内。",
        "en": "One FX100 serves 8 inference nodes; 100 GbE sustaining KV read-back for 8 nodes is a measured anchor. On the joint-test basis, storage spend stays within roughly 10% of the GPU budget."
      },
      "provenance": "estimated",
      "reports": [
        "R1",
        "R2",
        "R3"
      ]
    }
  ],
  "benchmark": {
    "report": "R2",
    "caption": {
      "zh": "Qwen3-Coder-480B-FP8 · TP=8 · 8 × MI308X · 长上下文冷恢复负载。TTFT 为 p50，单位秒；吞吐单位 tok/s；整段重算只测了并发 16 一档。",
      "en": "Qwen3-Coder-480B-FP8 · TP=8 · 8 × MI308X · long-context cold-recovery load. TTFT is p50 in seconds; throughput in tok/s; full recompute was measured at concurrency 16 only."
    },
    "units": {
      "ttft": "seconds (p50)",
      "throughput": "tokens per second"
    },
    "rows": [
      {
        "concurrency": 8,
        "fxTtft": 7.53,
        "fxTput": 56.6,
        "localTtft": 10.17,
        "localTput": 43.9,
        "recomputeTtft": null,
        "recomputeTput": null
      },
      {
        "concurrency": 16,
        "fxTtft": 11.85,
        "fxTput": 74.9,
        "localTtft": 17.31,
        "localTput": 53.6,
        "recomputeTtft": 149.48,
        "recomputeTput": 4.1
      },
      {
        "concurrency": 32,
        "fxTtft": 26.35,
        "fxTput": 71.6,
        "localTtft": 35.73,
        "localTput": 53.9,
        "recomputeTtft": null,
        "recomputeTput": null
      }
    ]
  },
  "platform": {
    "appliesTo": [
      "R1",
      "R2",
      "R3",
      "R4"
    ],
    "rows": [
      {
        "key": {
          "zh": "加速卡",
          "en": "Accelerators"
        },
        "value": {
          "zh": "8 × AMD Instinct MI308X（每卡 192 GB HBM，gfx942）",
          "en": "8 × AMD Instinct MI308X (192 GB HBM each, gfx942)"
        }
      },
      {
        "key": {
          "zh": "主机",
          "en": "Host"
        },
        "value": {
          "zh": "2 × AMD EPYC 9654（384 线程），内存约 1.5 TB",
          "en": "2 × AMD EPYC 9654 (384 threads), about 1.5 TB RAM"
        }
      },
      {
        "key": {
          "zh": "软件栈",
          "en": "Software stack"
        },
        "value": {
          "zh": "ROCm 7.2 · vLLM 0.20.1+rocm721 · LMCache（上游主线 2026-06-29 源码编译）",
          "en": "ROCm 7.2 · vLLM 0.20.1+rocm721 · LMCache (built from upstream mainline, 2026-06-29)"
        }
      },
      {
        "key": {
          "zh": "被测设备",
          "en": "Device under test"
        },
        "value": {
          "zh": "FX100 全闪 NVMe-oF 阵列：4 盘 RAID0（14 TB，XFS），RoCEv2，单口 100 GbE",
          "en": "FX100 all-flash NVMe-oF array: 4-drive RAID0 (14 TB, XFS), RoCEv2, one 100 GbE port"
        }
      },
      {
        "key": {
          "zh": "对照基线",
          "en": "Baselines"
        },
        "value": {
          "zh": "本地 NVMe 单盘（PCIe Gen4，2 TB）；无外存、整段重算",
          "en": "A single local NVMe drive (PCIe Gen4, 2 TB); no external storage, full recompute"
        }
      },
      {
        "key": {
          "zh": "480B 模型",
          "en": "480B model"
        },
        "value": {
          "zh": "Qwen3-Coder-480B-FP8（MoE，权重约 450 GB）",
          "en": "Qwen3-Coder-480B-FP8 (MoE, about 450 GB of weights)"
        }
      }
    ]
  },
  "products": [
    {
      "id": "fx100",
      "name": "FX100",
      "url": "https://vuke.top/products/fx100",
      "interface": "PCIe 3.0",
      "portGb": 100,
      "iopsMillion": 16,
      "aggregateTbps": null,
      "flash": "U.2",
      "available": true,
      "status": {
        "zh": "量产在售",
        "en": "In production"
      },
      "priceCny": {
        "platform": 160000,
        "perDrive": 8800,
        "driveCount": 24,
        "full": 371200,
        "rawTb": 184.3,
        "perTb": 2014,
        "note": {
          "zh": "满配整机参考价 = 平台 ¥160,000 + 24 × ¥8,800 = ¥371,200，裸容量 184.3 TB，约合 ¥2,014/TB。SSD 按参考市场价计、随行就市，正式价格以采购合同为准。",
          "en": "Fully populated reference price = platform ¥160,000 + 24 × ¥8,800 = ¥371,200 for 184.3 TB raw, about ¥2,014 per TB. SSDs are counted at a reference market price that moves with the market; the purchase contract sets the final price."
        }
      },
      "reports": [
        "R1",
        "R2",
        "R3",
        "R4",
        "R5",
        "R9"
      ]
    },
    {
      "id": "fx200",
      "name": "FX200",
      "url": "https://vuke.top/products/fx200",
      "interface": "PCIe 4.0",
      "portGb": 200,
      "iopsMillion": 32,
      "aggregateTbps": null,
      "flash": "U.2",
      "available": true,
      "status": {
        "zh": "量产在售",
        "en": "In production"
      },
      "priceCny": null,
      "reports": []
    },
    {
      "id": "fx300",
      "name": "FX300",
      "url": "https://vuke.top/products/fx300",
      "interface": "PCIe 5.0",
      "portGb": 400,
      "iopsMillion": 60,
      "aggregateTbps": null,
      "flash": "U.2",
      "available": true,
      "status": {
        "zh": "量产在售",
        "en": "In production"
      },
      "priceCny": null,
      "reports": []
    },
    {
      "id": "fx400",
      "name": "FX400",
      "url": "https://vuke.top/products/fx400",
      "interface": "PCIe 6.0",
      "portGb": 400,
      "iopsMillion": 140,
      "aggregateTbps": 4.8,
      "flash": "E1.S",
      "available": false,
      "status": {
        "zh": "测试机阶段，计划 2026 年底量产",
        "en": "Test units; volume production planned for late 2026"
      },
      "priceCny": null,
      "reports": []
    }
  ],
  "reports": [
    {
      "id": "R1",
      "date": "2026-07-03",
      "kind": {
        "zh": "性能测试报告",
        "en": "Benchmark report"
      },
      "platform": {
        "zh": "AMD Instinct MI308X × 8 · ROCm 7.2",
        "en": "8 × AMD Instinct MI308X · ROCm 7.2"
      },
      "title": {
        "zh": "FX100 大模型推理与训练综合性能测试报告（AMD MI308X × 8）",
        "en": "FX100 Comprehensive LLM Inference and Training Benchmark (8 × AMD MI308X)"
      },
      "summary": {
        "zh": "一份报告回答五个问题：LMCache 并行读补丁有多大作用、KV 分层在多卡下能否扩展、8 卡并发加载模型快多少、频繁切换模型时算力有效利用率如何变化、训练 Checkpoint 并发写能快多少。测试模型为 Qwen2.5-32B 与 7B。",
        "en": "One report, five questions: how much the LMCache parallel-read patch helps, whether KV tiering scales across GPUs, how much faster 8 GPUs load models concurrently, how effective utilization changes under frequent model switching, and how much faster concurrent training checkpoint writes become. Test models are Qwen2.5-32B and 7B."
      },
      "url": "https://vuke.top/evidence/R1",
      "urlEn": "https://vuke.top/en/evidence/R1"
    },
    {
      "id": "R2",
      "date": "2026-07-05",
      "kind": {
        "zh": "性能测试报告（正式版）",
        "en": "Benchmark report (official edition)"
      },
      "platform": {
        "zh": "AMD Instinct MI308X × 8 · ROCm 7.2",
        "en": "8 × AMD Instinct MI308X · ROCm 7.2"
      },
      "title": {
        "zh": "FX100 KV Cache 性能测试报告（480B · TP8 长上下文）",
        "en": "FX100 KV Cache Benchmark (480B, TP8, Long Context)"
      },
      "summary": {
        "zh": "Qwen3-Coder-480B-FP8 以八卡单实例（TP=8）的标准生产形态部署，在长上下文冷恢复负载下做三方对照：FX100、本地 NVMe 单盘、无外存整段重算，并按并发 8 / 16 / 32 做梯度。本站引用最多的吞吐与 TTFT 数字出自这份报告。",
        "en": "Qwen3-Coder-480B-FP8 deployed as one 8-GPU instance (TP=8), the standard production topology, under a long-context cold-recovery load. A three-way comparison of FX100, a single local NVMe drive and full recompute without external storage, stepped across concurrency 8 / 16 / 32. Most throughput and TTFT figures quoted on this site come from this report."
      },
      "url": "https://vuke.top/evidence/R2",
      "urlEn": "https://vuke.top/en/evidence/R2"
    },
    {
      "id": "R3",
      "date": "2026-07-06",
      "kind": {
        "zh": "性能测试汇总报告（正式版）",
        "en": "Benchmark summary report (official edition)"
      },
      "platform": {
        "zh": "AMD Instinct MI308X × 8 · ROCm 7.2",
        "en": "8 × AMD Instinct MI308X · ROCm 7.2"
      },
      "title": {
        "zh": "FX100 KV Cache 性能测试汇总报告（480B · TP4×2 · 全指标）",
        "en": "FX100 KV Cache Benchmark Summary (480B, TP4×2, All Metrics)"
      },
      "summary": {
        "zh": "同一台 8 卡服务器改为四卡双实例（TP4×2）部署，九档负载记录全部指标：TTFT 的 p50 / p90 / p99、TPOT、吞吐与盘带宽。这份报告还验证了两件事：fs:// 共享池能让两个实例热共享同一份缓存；同一档位做两轮独立测量，偏差不超过 5%。",
        "en": "The same 8-GPU server redeployed as two 4-GPU instances (TP4×2), with every metric recorded across nine load levels: TTFT p50 / p90 / p99, TPOT, throughput and disk bandwidth. The report also verifies two things: the fs:// shared pool lets both instances hot-share one cache, and two independent runs at the same level deviate by no more than 5%."
      },
      "url": "https://vuke.top/evidence/R3",
      "urlEn": "https://vuke.top/en/evidence/R3"
    },
    {
      "id": "R4",
      "date": "2026-07-06",
      "kind": {
        "zh": "第三方检测机构正式报告",
        "en": "Official third-party test report"
      },
      "platform": {
        "zh": "AMD Instinct MI308X × 8 · ROCm 7.2",
        "en": "8 × AMD Instinct MI308X · ROCm 7.2"
      },
      "title": {
        "zh": "FX100 KV Cache 性能测试报告（480B · 多实例形态 · 编号 -006）",
        "en": "FX100 KV Cache Benchmark (480B, Multi-Instance, No. -006)"
      },
      "summary": {
        "zh": "R2 与 R3 的正式汇编版，由第三方检测机构出具，报告编号 -006。如果你的采购或合规流程需要一份出自第三方机构的文件，引用这一份。",
        "en": "The official compilation of R2 and R3, issued by a third-party testing organization as report No. -006. If your procurement or compliance process needs a document from an independent body, this is the one to cite."
      },
      "url": "https://vuke.top/evidence/R4",
      "urlEn": "https://vuke.top/en/evidence/R4"
    },
    {
      "id": "R5",
      "date": "2026-07-03",
      "kind": {
        "zh": "性能测试报告（正式版，编号 -004）",
        "en": "Benchmark report (official edition, No. -004)"
      },
      "platform": {
        "zh": "沐曦 N260 单卡",
        "en": "Single MetaX N260 GPU"
      },
      "title": {
        "zh": "FX100 KV Cache 性能测试报告（14B · 显存效益）",
        "en": "FX100 KV Cache Benchmark (14B, HBM Efficiency)"
      },
      "summary": {
        "zh": "这份报告回答「外置 KV 层能不能当显存用」。14B 模型上做七组对照（GDS 直读、两档显存下的重算、低显存直读），用三步论证得出显存等效替代的结论。测试平台是沐曦 N260 单卡，方法论可以跨平台使用，但数字不与 MI308X 的结果混用。",
        "en": "This report answers whether an external KV tier can stand in for GPU memory. A seven-way comparison on a 14B model (GDS direct read, recompute at two HBM sizes, direct read with low HBM) builds a three-step argument for HBM equivalence. The platform is a single MetaX N260 GPU; the method carries across platforms, but the numbers are never mixed with MI308X results."
      },
      "url": "https://vuke.top/evidence/R5",
      "urlEn": "https://vuke.top/en/evidence/R5"
    },
    {
      "id": "R6",
      "date": "2026-07-07",
      "kind": {
        "zh": "部署与适配报告",
        "en": "Deployment and adaptation report"
      },
      "platform": {
        "zh": "AMD Instinct MI308X · ROCm 7.2",
        "en": "AMD Instinct MI308X · ROCm 7.2"
      },
      "title": {
        "zh": "ComfyUI + LTX-Video 2.3 全模型部署、适配与使用报告（V2）",
        "en": "ComfyUI + LTX-Video 2.3 Full-Model Deployment, Adaptation and Usage Report (V2)"
      },
      "summary": {
        "zh": "视频生成方向的适配证明：交付的 7 个模型（3 个 VAE、fp8 文本编码器、文本投影、2 个 LoRA）在 MI308X / ROCm 7.2 上全部实跑通过，没有出现 dtype、算子或内核错误。",
        "en": "Proof of adaptation for video generation: all 7 delivered models (3 VAEs, the fp8 text encoder, the text projection and 2 LoRAs) run end to end on MI308X / ROCm 7.2 with no dtype, operator or kernel errors."
      },
      "url": "https://vuke.top/evidence/R6",
      "urlEn": "https://vuke.top/en/evidence/R6"
    },
    {
      "id": "R7",
      "date": "2026-07-07",
      "kind": {
        "zh": "模型适配报告",
        "en": "Model adaptation report"
      },
      "platform": {
        "zh": "AMD Instinct MI308X",
        "en": "AMD Instinct MI308X"
      },
      "title": {
        "zh": "ComfyUI + LTX-Video 2.3 模型适配报告（AMD MI308X）",
        "en": "ComfyUI + LTX-Video 2.3 Model Adaptation Report (AMD MI308X)"
      },
      "summary": {
        "zh": "R6 的分模型细目：3 个 VAE 与 Singularity LoRA 逐个实跑的记录，以及在 MI308X 上需要注意的适配要点。",
        "en": "The per-model detail behind R6: individual run records for the 3 VAEs and the Singularity LoRA, plus the adaptation points that matter on MI308X."
      },
      "url": "https://vuke.top/evidence/R7",
      "urlEn": "https://vuke.top/en/evidence/R7"
    },
    {
      "id": "R8",
      "date": "2026-07",
      "kind": {
        "zh": "代码导出包与原始数据",
        "en": "Code export and raw data"
      },
      "platform": {
        "zh": "AMD Instinct MI308X · ROCm 7.2",
        "en": "AMD Instinct MI308X · ROCm 7.2"
      },
      "title": {
        "zh": "FX100 KV Cache 测试代码导出包与原始测试数据",
        "en": "FX100 KV Cache Test Code Export and Raw Benchmark Data"
      },
      "summary": {
        "zh": "让第三方能自己重跑一遍的全部材料：LMCache 并行读补丁（git patch 与修改前后全文）、负载客户端、编排脚本、探针取证脚本，以及原始测试数据。",
        "en": "Everything a third party needs to rerun the tests: the LMCache parallel-read patch (git patch plus full before/after sources), the load client, orchestration scripts, forensic probe scripts and the raw benchmark data."
      },
      "url": "https://vuke.top/evidence/R8",
      "urlEn": "https://vuke.top/en/evidence/R8"
    },
    {
      "id": "R9",
      "date": "2026-05-30",
      "kind": {
        "zh": "性能测试报告",
        "en": "Benchmark report"
      },
      "platform": {
        "zh": "华为 Atlas 910B × 8（Kunpeng-920）",
        "en": "8 × Huawei Atlas 910B (Kunpeng-920)"
      },
      "title": {
        "zh": "FX100 与华为昇腾 910B 模型推理与训练性能测试（对照 NFS 基线）",
        "en": "FX100 on Huawei Ascend 910B: Inference and Training Benchmark (vs. NFS Baseline)"
      },
      "summary": {
        "zh": "在华为 Atlas 910B × 8 的昇腾平台上，以 NFS 为基线做三组对照：模型推理加载（DeepSeek-32B / 70B）、训练权重与 Checkpoint 存取（Qwen-7B）、训练数据加速（YOLOv8 / COCO）。平台是昇腾 910B 而不是 MI308X，本站引用时都会注明。",
        "en": "Three comparison groups against an NFS baseline on an Ascend platform with 8 × Huawei Atlas 910B: model loading for inference (DeepSeek-32B / 70B), training weight and checkpoint I/O (Qwen-7B), and training-data acceleration (YOLOv8 / COCO). The platform is Ascend 910B, not MI308X, and is labeled as such wherever this site quotes it."
      },
      "url": "https://vuke.top/evidence/R9",
      "urlEn": "https://vuke.top/en/evidence/R9"
    }
  ],
  "engagement": {
    "headline": {
      "zh": "先联测，后决策",
      "en": "Test jointly first, decide afterwards"
    },
    "summary": {
      "zh": "合作从一次约 10 周的门禁化联测开始。四道门禁逐级放行，任何一道不达标就在那里停下，不必为没有被证实的收益买单。",
      "en": "Every engagement starts with a gated joint test of about 10 weeks. Four gates are passed in sequence; if one is missed, the project stops there, so you never pay for a benefit that was not demonstrated."
    },
    "gates": [
      {
        "id": "G1",
        "week": {
          "zh": "第 2 周",
          "en": "Week 2"
        },
        "name": {
          "zh": "到货验收",
          "en": "Delivery acceptance"
        },
        "desc": {
          "zh": "物料全部到货并点验合格；每台阵列上电自检通过，24 块盘全部识别。",
          "en": "All materials delivered and inspected; every array passes its power-on self-test with all 24 drives recognized."
        }
      },
      {
        "id": "G2",
        "week": {
          "zh": "第 4 周",
          "en": "Week 4"
        },
        "name": {
          "zh": "单机基线",
          "en": "Single-node baseline"
        },
        "desc": {
          "zh": "复现单节点 8 卡推理基线，与 R1 口径的偏差在 ±10% 以内；单阵列本地压测达到实测锚点 83% 的线效。",
          "en": "The single-node 8-GPU inference baseline is reproduced within ±10% of the R1 figures; a local stress test of one array reaches 83% line efficiency against the measured anchor."
        }
      },
      {
        "id": "G3",
        "week": {
          "zh": "第 7 周",
          "en": "Week 7"
        },
        "name": {
          "zh": "联测主门禁",
          "en": "Main joint-test gate"
        },
        "desc": {
          "zh": "KV 分层加速须同时满足：TTFT 降幅 ≥25%，吞吐提升落在 +29–40% 的实测区间内；8 个节点共享一台阵列时，回读带宽逼近网卡上限。",
          "en": "KV tiering must deliver both a TTFT reduction of ≥25% and a throughput gain inside the measured +29–40% band; with 8 nodes sharing one array, read-back bandwidth approaches the NIC ceiling."
        }
      },
      {
        "id": "G4",
        "week": {
          "zh": "第 10 周",
          "en": "Week 10"
        },
        "name": {
          "zh": "稳定性与验收",
          "en": "Stability and sign-off"
        },
        "desc": {
          "zh": "72 小时连续压测不中断；注入单盘、单链路故障时业务无感；出具签字版联测报告。",
          "en": "72 hours of continuous load without interruption; single-drive and single-link fault injection stays invisible to the workload; a signed joint-test report is issued."
        }
      }
    ],
    "note": {
      "zh": "完整测算模型在签署 NDA 后以 Python 源码提供，你可以换成自己的参数重新计算。",
      "en": "The full costing model is provided as Python source after an NDA, so you can rerun it with your own parameters."
    }
  }
}