dc-thermal
Menu菜单

Products产品 · GPU

GPU roadmapGPU 路线图

Chip and rack power ratings: design heat load, not measured load芯片与机柜额定功率：设计热负荷，不是实测负载
202420252026202720282029 Chip or module power, W芯片或模块功率，W05001,0001,5002,000H200 SXM 700 WH200 SXM 700 WB200 1,000 WB200 1,000 WGB300 GPU 1,400 WGB300 GPU 1,400 WMI300X 750 WMI300X 750 WMI325X 1,000 WMI325X 1,000 WMI355X 1,400 WMI355X 1,400 WMaia 200 750 WMaia 200 750 WMTIA 300 800 WMTIA 300 800 WMTIA 400 1,200 WMTIA 400 1,200 WMTIA 450 1,400 WMTIA 450 1,400 WMTIA 500 1,700 WMTIA 500 1,700 WGaudi 3 HL-325L 900 WGaudi 3 HL-325L 900 W Rack or server power, kW (NVIDIA)机柜或服务器功率，kW（NVIDIA）0100200300400GB200 NVL72 120 kWGB200 NVL72 120 kWGB300 NVL72 135 kWGB300 NVL72 135 kWVera Rubin NVL72 330 kW (DSX design envelope)Vera Rubin NVL72 330 kW（DSX 设计包络）DGX H200 10.2 kWDGX H200 10.2 kWDGX B200 14.3 kWDGX B200 14.3 kWDGX B300 15 kWDGX B300 15 kW

Chip or module power芯片或模块功率

  • H200 SXM 700 W GPU TDP, up toGPU TDP，最高
  • MI300X 750 W OAM module, maximum TBPOAM 模块，最大 TBP
  • B200 (DGX B200) 1,000 W GPU TGP, maximum (default cap 700 W)GPU TGP，最大值（默认上限 700 W）
  • Gaudi 3 HL-325L 900 W OAM module TDPOAM 模块 TDP
  • MI325X 1,000 W OAM module TBP, peakOAM 模块 TBP，峰值
  • GB300 GPU 1,400 W GPU maximum in the NVL72 rack budgetNVL72 机柜功率预算中的单 GPU 最大值
  • MI355X 1,400 W OAM module TBPOAM 模块 TBP
  • Maia 200 750 W SoC TDP envelopeSoC TDP 上限
  • MTIA 300 800 W Module TDP模块 TDP
  • MTIA 400 1,200 W Module TDP模块 TDP
  • MTIA 450 1,400 W Module TDP, planned模块 TDP，计划
  • MTIA 500 1,700 W Module TDP, planned模块 TDP，计划

Rack or server power (NVIDIA)机柜或服务器功率（NVIDIA）

  • DGX H200 10.2 kW 8-GPU server, maximum system power (AC input)8 GPU 服务器，系统最大功率（交流输入）
  • GB200 NVL72 120 kW Designed rack power (Mission Control); the DGX GB user guide gives about 120 kW for the GB rack family设计机柜功率（Mission Control）；DGX GB 用户指南对 GB 机柜系列给出约 120 kW
  • DGX B200 14.3 kW 8-GPU server, maximum system power (AC input)8 GPU 服务器，系统最大功率（交流输入）
  • GB300 NVL72 135 kW Designed rack power (NVIDIA's reference architecture gives up to 142 kW for the full rack)设计机柜功率（NVIDIA 参考架构给出整柜最高 142 kW）
  • DGX B300 15 kW 8-GPU server, maximum system power (AC input); NVIDIA also gives about 14 kW on the product page and 14.5 kW power consumption in the user guide8 GPU 服务器，系统最大功率（交流输入）；NVIDIA 产品页另写约 14 kW，用户指南另列功耗 14.5 kW
  • Vera Rubin NVL72 (DSX) 330 kW Cabinet TDP of the DSX facility reference design (a design envelope)DSX 设施参考设计的机柜 TDP（设计包络）

Shapes group similar power boundaries: 形状按相近的功率口径分组：GPU TDP, TGP or per-GPU budget maximum (NVIDIA)GPU TDP、TGP 或单 GPU 预算最大值（NVIDIA） OAM module board power (AMD) or TDP (Intel)OAM 模块板卡功率（AMD）或 TDP（Intel） SoC or module TDP (Microsoft, Meta)SoC 或模块 TDP（Microsoft、Meta） designed rack power (GB200, GB300)设计机柜功率（GB200、GB300） 8-GPU server, maximum system power, AC input (DGX H200, B200, B300)8 GPU 服务器，系统最大功率，交流输入（DGX H200、B200、B300） cabinet TDP of a facility reference design (Vera Rubin)设施参考设计的机柜 TDP（Vera Rubin）

Filled: disclosed specification. Hollow: official plan. No line joins the points because their boundaries differ. Each point is placed at the product's announcement or launch date; the value is today's official specification. Meta MTIA 450 and 500 sit at the start of their planned deployment year (early 2027; 2027), not at an announcement date. Hover or focus a point for details.实心：已公布规格。空心：官方计划。各点口径不同，所以不连线。每个点放在产品发布或上市日期；数值取当前官方规格。Meta MTIA 450 和 500 放在计划部署年份的起点（2027 年初；2027 年），不是发布日期。悬停或聚焦数据点可查看详情。
Official sources checked 2026-10-07; rack note rechecked 2026-10-17. Boundaries differ (GPU TDP or TGP, OAM board power, SoC or module TDP; rack power, 8-GPU server maximum system power or cabinet TDP), so read the direction, not exact ratios between points; none is measured load. DGX H200, B200 and B300 are single 8-GPU servers at the maximum system power (AC input) in NVIDIA's user guides (the 10.2 kW row is shared by DGX H100 and H200); they sit at the date of the HGX platform announcement, which names the 8-GPU boards rather than the DGX systems. For cooling, the same guides give heat output of 38,557 BTU/hr (about 11.3 kW) for DGX H200, 48,794 BTU/hr (14.3 kW) for B200 and 49,476 BTU/hr (14.5 kW) for B300. How many servers go in a rack is a design choice: NVIDIA's DGX SuperPOD references put two DGX H200 or B200 systems in a rack, and four DGX B300 in a rack at about 56 kW. For DGX B300, NVIDIA also gives about 14 kW (product page) and 14.5 kW power consumption (user guide); the chart uses the guide's 15 kW system maximum, the same basis as H200 and B200. GB200 NVL72 120 kW and GB300 NVL72 135 kW are the designed rack power in the same NVIDIA Mission Control table. NVIDIA's GB300 reference architecture also says a full rack requires "up to 142 kW", a different boundary. The Vera Rubin 330 kW is the cabinet TDP of NVIDIA's DSX facility reference design, not a rack rating comparable with the GB200 and GB300 rack power. In a separate NVIDIA test on inference workloads (different models), provisioned rack power before NVIDIA's MaxLPS power management was 136 kW for Vera Rubin NVL72 and 125 kW for GB200 NVL72. Do not read 135 → 330 kW as growth between generations. We found no official absolute chip or rack power for Google, AWS or Huawei, and no verified official value yet for the Rubin GPU, the AMD Helios rack or later NVIDIA racks; missing is not zero.官方来源核查 2026-10-07；机柜说明于 2026-10-17 复核。各点口径不同（GPU TDP 或 TGP、OAM 板卡功率、SoC 或模块 TDP；机柜功率、8 GPU 服务器系统最大功率或机柜 TDP），只宜看方向，不宜计算点与点之间的精确倍数；都不是实测负载。DGX H200、B200 和 B300 是单台 8 GPU 服务器，数值取 NVIDIA 用户指南中的系统最大功率（交流输入），其中 10.2 kW 为 DGX H100 与 H200 共用；它们放在 HGX 平台的发布日期，该发布提到的是 8 GPU 主板而不是 DGX 整机。用于冷却设计时，同一批指南给出的散热量为：DGX H200 38,557 BTU/hr（约 11.3 kW），B200 48,794 BTU/hr（14.3 kW），B300 49,476 BTU/hr（14.5 kW）。一个机柜放几台由设计决定：NVIDIA DGX SuperPOD 参考设计每柜放 2 台 DGX H200 或 B200，DGX B300 每柜放 4 台、约 56 kW。DGX B300 另有两个 NVIDIA 数值：产品页约 14 kW，用户指南功耗 14.5 kW；图中取指南的 15 kW 系统最大值，与 H200、B200 口径相同。GB200 NVL72 的 120 kW 和 GB300 NVL72 的 135 kW 取自 NVIDIA Mission Control 同一张表中的设计机柜功率。NVIDIA GB300 参考架构另写整柜需求“最高 142 kW”，口径不同。Vera Rubin 的 330 kW 是 NVIDIA DSX 设施参考设计的机柜 TDP，不能与 GB200、GB300 的机柜功率直接比较。NVIDIA 另一项推理负载测试（所用模型不同）中，启用 MaxLPS 功率管理前的配置机柜功率为 Vera Rubin NVL72 136 kW、GB200 NVL72 125 kW。不要把 135 → 330 kW 理解为代际增长。未找到 Google、AWS 和华为官方公布的芯片或机柜绝对功率；Rubin GPU、AMD Helios 机柜和 NVIDIA 后续机柜暂无经核实的官方数值；缺失不等于零。
Drawn from the evidence base each time the site is built, so a newly adopted event appears on the next build. The quarterly roadmap review adds and rechecks events.每次构建网站时从证据底座生成，新采纳的事件在下次构建时出现。季度路线图复核负责新增和复查事件。
Chip roadmap: AI chips and platforms from eight vendors, 2023 Q4 to 2029 Q3芯片路线图：八家厂商的 AI 芯片与平台，2023 年第四季度至 2029 年第三季度
202420252026202720282029 Checked 2026-10-12核查 2026-10-12 NVIDIAH200H200Blackwell GB200Blackwell GB200Blackwell Ultra GB300Blackwell Ultra GB300Vera RubinVera RubinVera Rubin NVL72Vera Rubin NVL72Rubin partner productsRubin partner products800 V DC racks800 V DC racksRubin UltraRubin UltraFeynmanFeynman AMDMI300XMI300XMI325XMI325XMI355XMI355XMI455X / HeliosMI455X / HeliosMI500MI500MI600MI600 GoogleTrillium (TPU v6)Trillium (TPU v6)Ironwood (TPU v7)Ironwood (TPU v7)TPU 8t / 8iTPU 8t / 8i AWSTrainium2Trainium2Trainium3Trainium3Trainium4Trainium4 MicrosoftMaia 100Maia 100Maia 200Maia 200 MetaMTIA 300MTIA 300MTIA 400MTIA 400MTIA 450MTIA 450MTIA 500MTIA 500 HuaweiAtlas 950 SuperPoDAtlas 950 SuperPoDAscend 960DTAscend 960DTAscend 960PRAscend 960PRAscend 970Ascend 970Ascend 980Ascend 980 IntelGaudi 3Gaudi 3Crescent IslandCrescent Island
  1. NVIDIA
    • H200H200
    • Blackwell GB200Blackwell GB200
    • Blackwell Ultra GB300Blackwell Ultra GB300
    • Vera RubinVera Rubin
    • Vera Rubin NVL72Vera Rubin NVL72 (plan)（计划）
    • Rubin partner productsRubin partner products (plan)（计划）
    • 800 V DC racks800 V DC racks (plan)（计划）
    • Rubin UltraRubin Ultra (plan)（计划）
    • FeynmanFeynman (plan)（计划）
  2. AMD
    • MI300XMI300X
    • MI325XMI325X
    • MI355XMI355X
    • MI455X / HeliosMI455X / Helios
    • MI500MI500 (plan)（计划）
    • MI600MI600 (plan)（计划）
  3. Google
    • Trillium (TPU v6)Trillium (TPU v6)
    • Ironwood (TPU v7)Ironwood (TPU v7)
    • TPU 8t / 8iTPU 8t / 8i
  4. AWS
    • Trainium2Trainium2
    • Trainium3Trainium3
    • Trainium4Trainium4 (plan)（计划）
  5. Microsoft
    • Maia 100Maia 100
    • Maia 200Maia 200
  6. Meta
    • MTIA 300MTIA 300
    • MTIA 400MTIA 400
    • MTIA 450MTIA 450 (plan)（计划）
    • MTIA 500MTIA 500 (plan)（计划）
  7. Huawei
    • Atlas 950 SuperPoDAtlas 950 SuperPoD
    • Ascend 960DTAscend 960DT (plan)（计划）
    • Ascend 960PRAscend 960PR (plan)（计划）
    • Ascend 970Ascend 970 (plan)（计划）
    • Ascend 980Ascend 980 (plan)（计划）
  8. Intel
    • Gaudi 3Gaudi 3
    • Crescent IslandCrescent Island (plan)（计划）
Solid capsule: happened (announcement, availability or production, as the event says). Dashed capsule: a stated plan, placed at the start of its window. Hover or focus a capsule for the event and its date precision.实线胶囊：已发生（发布、供货或量产，以事件描述为准）。虚线胶囊：公司公布的计划，放在其时间窗口起点。悬停或聚焦胶囊可查看事件及日期精度。
34 events, sources checked 2026-10-07 to 2026-10-12. Where a vendor publishes a rack configuration, such as NVIDIA's Vera Rubin NVL72, its rack power and coolant temperature are the design point for the thermal equipment (see Products); most entries here publish neither.34 个事件，来源核查 2026-10-07 至 2026-10-12。厂商公布机柜配置时（例如 NVIDIA Vera Rubin NVL72），其机柜功率和冷却液温度就是热管理设备的设计点（见产品页）；这里多数条目两者都未公布。