Documentation

Performance Specifications

This section describes only the performance specifications of Ascend AI Processors of each product, which may differ based on the memory and CPU configurations of the respective product. For details about the product performance specifications, see the official descriptions at the Ascend community.

This section uses a single Ascend AI Processor as an example to describe the performance indicators in basic scenarios. The performance indicators of other resolutions can be estimated in the same way.

Performance Specifications (Atlas 200/300/500 Inference Product)

The following performance data is based on executing acl.rt.synchronize_stream once after 10 asynchronous media data processing tasks are delivered to a stream.

  • For non-8K images in the single-device scenario, the performance specifications are as follows (one channel corresponds to one thread):

    For 1080p images, if image data is transferred from the host to the device, the maximum total frame rate is about 1000 FPS due to the limited transfer bandwidth.

    For 4K images, if image data is transferred from the host to the device, the maximum total frame rate is about 250 FPS due to the limited transfer bandwidth.
    Table 1 Performance specifications of the Ascend AI Processor on the Atlas 200/300/500 Inference Product

    Example

    Total Frame Rate
    • Input resolution: 1080p (1920 x 1080)
    • Output resolution: 1080p (1920 x 1080)
    • Input/Output format: YUV420SP
    • n channels (n < 4)

    n * 360fps
    • Input resolution: 1080p (1920 x 1080)
    • Output resolution: 1080p (1920 x 1080)
    • Input/Output format: YUV420SP
    • n channels (n ≥ 4)

    1440 FPS
    • Input resolution: 4K (3840 x 2160)
    • Output resolution: 4K (3840 x 2160)
    • Input/Output format: YUV420SP
    • n channels (n < 4)

    n * 90fps
    • Input resolution: 4K (3840 x 2160)
    • Output resolution: 4K (3840 x 2160)
    • Input/Output format: YUV420SP
    • n channels (n ≥ 4)

    360 FPS
  • For 8K images in the single-device scenario, the performance specifications are as follows (one channel corresponds to one thread).
    Table 2 Performance specifications of the Ascend AI Processor on the Atlas 200/300/500 Inference Product

    Scenario

    Total Frame Rate
    • Input resolution: 8K (7680 x 4320)
    • Output resolution: 1080p (1920 x 1080)
    • Input/Output format: YUV420SP
    • n channels (n < 4)

    n * 25fps
    • Input resolution: 8K (7680 x 4320)
    • Output resolution: 1080p (1920 x 1080)
    • Input/Output format: YUV420SP
    • n channels (n ≥ 4)

    100 FPS
    • Input resolution: 8K (7680 x 4320)
    • Output resolution: 4K (3840 x 2160)
    • Input/Output format: YUV420SP
    • n channels (n < 4)

    n * 25fps
    • Input resolution: 8K (7680 x 4320)
    • Output resolution: 4K (3840 x 2160)
    • Input/Output format: YUV420SP
    • n channels (n ≥ 4)

    100 FPS
    • The VPC processing performance is closely related to the image resolution during processing. The maximum resolution of the input and output images is used as the reference resolution. A larger reference resolution indicates a longer processing time and lower performance.
    • A VPC batch processing API (whose name contains "batch", for example, acl.media.dvpp_vpc_batch_crop_async) allows the parallel image processing on multiple VPC hardware units in the DVPP. Therefore, a single channel can reach the maximum total frame rate.

Performance Specifications (Atlas Training Series Product)

The following performance data is based on executing acl.rt.synchronize_stream once after 10 asynchronous media data processing tasks are delivered to a stream.

  • For non-8K images in the single-device scenario, the performance specifications are as follows (one channel corresponds to one thread).
    Table 3 Performance specifications of the Ascend AI Processor on the Atlas Training Series Product

    Example

    Total Frame Rate
    • Input resolution: 1080p (1920 x 1080)
    • Output resolution: 1080p (1920 x 1080)
    • Input/Output format: YUV420SP
    • n channels (n < 16)

    n * 360fps
    • Input resolution: 1080p (1920 x 1080)
    • Output resolution: 1080p (1920 x 1080)
    • Input/Output format: YUV420SP
    • n channels (n ≥ 16)

    5760 FPS
    • Input resolution: 4K (3840 x 2160)
    • Output resolution: 4K (3840 x 2160)
    • Input/Output format: YUV420SP
    • n channels (n < 16)

    n * 90fps
    • Input resolution: 4K (3840 x 2160)
    • Output resolution: 4K (3840 x 2160)
    • Input/Output format: YUV420SP
    • n channels (n ≥ 16)

    1440 FPS
  • For 8K images in the single-device scenario, the performance specifications are as follows (one channel corresponds to one thread).
    Table 4 Performance specifications of the Ascend AI Processor on the Atlas Training Series Product

    Scenario

    Total Frame Rate
    • Input resolution: 8K (7680 x 4320)
    • Output resolution: 1080p (1920 x 1080)
    • Input/Output format: YUV420SP
    • n channels (n < 16)

    n * 25fps
    • Input resolution: 8K (7680 x 4320)
    • Output resolution: 1080p (1920 x 1080)
    • Input/Output format: YUV420SP
    • n channels (n ≥ 16)

    400 FPS
    • Input resolution: 8K (7680 x 4320)
    • Output resolution: 4K (3840 x 2160)
    • Input/Output format: YUV420SP
    • n channels (n < 16)

    n * 25fps
    • Input resolution: 8K (7680 x 4320)
    • Output resolution: 4K (3840 x 2160)
    • Input/Output format: YUV420SP
    • n channels (n ≥ 16)

    400 FPS

    A VPC batch processing API (whose name contains "batch", for example, acl.media.dvpp_vpc_batch_crop_async) allows the parallel image processing on multiple VPC hardware units in the DVPP. Therefore, a single channel can reach the maximum total frame rate.

Parent topic: VPC