Virtualization Templates

Table 1 shows virtual instance templates currently supported by each product.

Table 1 Virtual instance templates

Product Model	Virtual Instance Template	Description
Atlas training series products (30 or 32 AICores)	Virtual instance templates include: vir02, vir04, vir08, vir16.	The number after vir indicates the number of AICores.
Atlas inference series products (8 AICores)	Virtual instance templates include: vir01, vir02, vir04, vir02_1c, vir04_3c, vir04_3c_ndvpp, vir04_4c_dvpp.	The number after vir indicates the number of AICores. The number before c indicates the number of AICPUs. dvpp indicates that all digital visual preprocessing modules (i.e., VPC, VDEC, JPEGD, PNGD, VENC, JPEGE) are included during virtualization. ndvpp indicates that there are no digital visual preprocessing hardware resources during virtualization.
Atlas A2 training series products (24 AICores)	Virtual instance templates include: vir06_1c_16g, vir12_3c_32g.	The number after vir indicates the number of AICores. The number before c indicates the number of AICPUs. The number before g indicates the memory size.
Atlas A2 inference series products (20 AICores)	Virtual instance templates include: vir05_1c_8g, vir10_3c_16g_nm, vir10_4c_16g_m, vir10_3c_16g, vir10_3c_32g, vir05_1c_16g.	The number after vir indicates the number of AICores. The number before c indicates the number of AICPUs. m, same as dvpp, indicates that all digital visual preprocessing modules (i.e., VPC, VDEC, JPEGD, PNGD, VENC, JPEGE) are included during virtualization. nm, same as ndvpp, indicates that there are no digital visual preprocessing hardware resources during virtualization. The number before g indicates the memory size.
Atlas A3 training series products (48 AICores)	Virtual instance templates include: vir06_1c_16g, vir12_3c_32g.	The number after vir indicates the number of AICores. The number before c indicates the number of AICPUs. The number before g indicates the memory size.
Atlas A3 inference series products (40 AICores)	Virtual instance templates include: vir05_1c_16g, vir10_3c_32g.	The number after vir indicates the number of AICores. The number before c indicates the number of AICPUs. The number before g indicates the memory size.
NOTE: The templates supported by a specific server can be queried using the npu-smi info -t template-info command.

The Ascend AI processor includes hardware resources such as AICore, AICPU, DVPP, and memory. Their main purposes are as follows:

AICore is mainly used for matrix multiplication and other computations, and is suitable for convolutional models.
AICPU is primarily responsible for executing CPU-type operators (including control operators, scalars, vectors, and other general-purpose computations).
Virtual instances (creating vNPUs for specified chips) enable SRIOV, converting data CPUs into AICPUs. As a result, the number of AICPUs displayed in the NPU information changes.
DVPP (Digital Vision Pre‑Processing) is a module that provides pre‑processing capabilities for video and image data in specific formats, including decoding, scaling, and encoding of processed video and images. It comprises the following modules.
- VPC (Vision Pre‑processing Core): Provides capabilities such as image scaling, color space conversion, bit‑depth reduction, format conversion, and block‑based cropping/transformation.
- VDEC (Video Decoder): Provides decoding capabilities for video in specific formats
- JPEGD (JPEG Decoder): Provides decoding capabilities for images in JPEG format.
- PNGD (PNG Decoder): Provides decoding capabilities for images in PNG format.
- VENC (Video Encoder): Provides encoding capabilities for video in specific formats.
- JPEGE (JPEG Encoder): Provides the ability to encode images and output them in JPEG format.