你当前正在访问 Microsoft Azure Global Edition 技术文档网站。 如果需要访问由世纪互联运营的 Microsoft Azure 中国技术文档网站,请访问 https://docs.azure.cn

托管联机终结点 SKU 列表

下表显示了 Azure 机器学习托管联机终结点支持的虚拟机 (VM) 库存单位 (SKU)。 每个 SKU 都是分配给可购买的特定 VM 的唯一字母数字代码。

  • 可以将表中列出的完整 SKU 名称用于 Azure CLI 或 Azure 资源管理器模板(ARM 模板)请求来创建和更新部署。

  • 有关 CPU 和 RAM 等配置详细信息,请参阅 Azure 机器学习定价VM 大小

家族名称 VM 大小名称 支持 Infiniband 体系结构 numberOfGPU numberOfCores 跳过 20% 预留
standardDASv4Family STANDARD_D2AS_V4 - Cpu 0 2 -
standardDASv4Family STANDARD_D4AS_V4 - Cpu 0 4 -
standardDASv4Family STANDARD_D8AS_V4 - Cpu 0 8 -
standardDASv4Family STANDARD_D16AS_V4 - Cpu 0 16 -
standardDASv4Family STANDARD_D32AS_V4 - Cpu 0 32 -
standardDASv4Family STANDARD_D48AS_V4 - Cpu 0 48 -
standardDASv4Family STANDARD_D64AS_V4 - Cpu 0 64 -
standardDASv4Family STANDARD_D96AS_V4 - Cpu 0 96 -
standardDAv4Family STANDARD_D2A_V4 - Cpu 0 2 -
standardDAv4Family STANDARD_D4A_V4 - Cpu 0 4 -
standardDAv4Family STANDARD_D8A_V4 - Cpu 0 8 -
standardDAv4Family STANDARD_D16A_V4 - Cpu 0 16 -
standardDAv4Family STANDARD_D32A_V4 - Cpu 0 32 -
standardDAv4Family STANDARD_D48A_V4 - Cpu 0 48 -
standardDAv4Family STANDARD_D64A_V4 - Cpu 0 64 -
standardDAv4Family STANDARD_D96A_V4 - Cpu 0 96 -
standardDSv2Family STANDARD_DS1_V2 - Cpu 0 1 -
standardDSv2Family STANDARD_DS2_V2 - Cpu 0 2 -
standardDSv2Family STANDARD_DS3_V2 - Cpu 0 4 -
standardDSv2Family STANDARD_DS4_V2 - Cpu 0 8 -
standardDSv2Family STANDARD_DS5_V2 - Cpu 0 16 -
standardESv3Family STANDARD_E2S_V3 - Cpu 0 2 -
standardESv3Family STANDARD_E4S_V3 - Cpu 0 4 -
standardESv3Family STANDARD_E8S_V3 - Cpu 0 8 -
standardESv3Family STANDARD_E16S_V3 - Cpu 0 16 -
standardESv3Family STANDARD_E32S_V3 - Cpu 0 32 -
standardESv3Family STANDARD_E48S_V3 - Cpu 0 48 -
standardESv3Family STANDARD_E64S_V3 - Cpu 0 64 -
standardFSv2Family STANDARD_F2S_V2 - Cpu 0 2 -
standardFSv2Family STANDARD_F4S_V2 - Cpu 0 4 -
standardFSv2Family STANDARD_F8S_V2 - Cpu 0 8 -
standardFSv2Family STANDARD_F16S_V2 - Cpu 0 16 -
standardFSv2Family STANDARD_F32S_V2 - Cpu 0 32 -
standardFSv2Family STANDARD_F48S_V2 - Cpu 0 48 -
standardFSv2Family STANDARD_F64S_V2 - Cpu 0 64 -
standardFSv2Family STANDARD_F72S_V2 - Cpu 0 72 -
standardFXMDVSFamily STANDARD_FX4MDS - Cpu 0 4 -
standardFXMDVSFamily STANDARD_FX12MDS - Cpu 0 12 -
standardFXMDVSFamily STANDARD_FX24MDS - Cpu 0 24 -
standardFXMDVSFamily STANDARD_FX36MDS - Cpu 0 36 -
standardFXMDVSFamily STANDARD_FX48MDS - Cpu 0 48 -
standardLASv3Family STANDARD_L8AS_V3 - Cpu 0 8 -
standardLASv3Family STANDARD_L16AS_V3 - Cpu 0 16 -
standardLASv3Family STANDARD_L32AS_V3 - Cpu 0 32 -
standardLASv3Family STANDARD_L48AS_V3 - Cpu 0 48 -
standardLASv3Family STANDARD_L64AS_V3 - Cpu 0 64 -
standardLASv3Family STANDARD_L80AS_V3 - Cpu 0 80 -
standardLSv2Family STANDARD_L8S_V2 - Cpu 0 8 -
standardLSv2Family STANDARD_L16S_V2 - Cpu 0 16 -
standardLSv2Family STANDARD_L32S_V2 - Cpu 0 32 -
standardLSv2Family STANDARD_L48S_V2 - Cpu 0 48 -
standardLSv2Family STANDARD_L64S_V2 - Cpu 0 64 -
standardLSv2Family STANDARD_L80S_V2 - Cpu 0 80 -
standardLSv3Family STANDARD_L8S_V3 - Cpu 0 8 -
standardLSv3Family STANDARD_L16S_V3 - Cpu 0 16 -
standardLSv3Family STANDARD_L32S_V3 - Cpu 0 32 -
standardLSv3Family STANDARD_L48S_V3 - Cpu 0 48 -
standardLSv3Family STANDARD_L64S_V3 - Cpu 0 64 -
standardLSv3Family STANDARD_L80S_V3 - Cpu 0 80 -
standardNCADSA100v4Family STANDARD_NC24ADS_A100_V4 - NvidiaGpu 1 24
standardNCADSA100v4Family STANDARD_NC48ADS_A100_V4 - NvidiaGpu 2 48
standardNCADSA100v4Family STANDARD_NC96ADS_A100_V4 - NvidiaGpu 4 96
标准NCASv3_T4系列 STANDARD_NC4AS_T4_V3 - NvidiaGpu 1 4 -
标准NCASv3_T4系列 STANDARD_NC8AS_T4_V3 - NvidiaGpu 1 8 -
标准NCASv3_T4系列 STANDARD_NC16AS_T4_V3 - NvidiaGpu 1 16 -
标准NCASv3_T4系列 STANDARD_NC64AS_T4_V3 - NvidiaGpu 4 64 -
standardNCSv2Family STANDARD_NC6S_V2 - NvidiaGpu 1 6 -
standardNCSv2Family STANDARD_NC12S_V2 - NvidiaGpu 2 12 -
standardNCSv2Family STANDARD_NC24S_V2 - NvidiaGpu 4 24 -
standardNCSv3Family STANDARD_NC6S_V3 - NvidiaGpu 1 6 -
standardNCSv3Family STANDARD_NC12S_V3 - NvidiaGpu 2 12 -
standardNCSv3Family STANDARD_NC24S_V3 - NvidiaGpu 4 24 -
standardNCADSH100v5Family STANDARD_NC40ADS_H100_V5 - NvidiaGpu 1 40
standardNCADSH100v5Family STANDARD_NC80ADIS_H100_V5 - NvidiaGpu 2 80
标准NDAMSv4_A100Family STANDARD_ND96AMSR_A100_V4 NvidiaGpu 8 96
标准NDASv4_A100系列 STANDARD_ND96ASR_V4 NvidiaGpu 8 96
standardNDSv2Family STANDARD_ND40RS_V2 NvidiaGpu 8 40
standardNDv5H100Family STANDARD_ND96IS_H100_v5 - NvidiaGpu 8 96
standardNDv5H100Family STANDARD_ND96ISR_H100_v5 NvidiaGpu 8 96

注意

小型 VM SKU(例如Standard_DS1_v2Standard_F2s_v2,对于较大的模型而言可能太小),并可能导致容器终止,因为内存不足、磁盘空间不足或探测失败,因为启动容器需要太长的时间。 如果遇到 OutOfQuota 错误ReourceNotReady 错误,请尝试更大的 VM SKU。 如果要降低使用托管在线终端节点部署多个模型的成本,请参阅多个本地模型的部署

注意

我们建议在生产方案中部署 3 个以上的实例。 此外,Azure 机器学习会预留 20% 的计算资源,用于对某些 VM SKU 执行升级,如用于部署的虚拟机配额分配中所述。 在“跳过 20% 预留”列中指定免除此额外配额预留的 VM SKU。