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 takes a single Ascend AI Processor as an example to describe the performance specifications in basic scenarios, as shown in the table below. The performance at other resolutions can be estimated in the same way.

In the following table, one channel corresponds to one thread, 720p indicates a resolution of 1280 x 720 pixels, 1080p indicates a resolution of 1920 x 1080 pixels, and 4K indicates a resolution of 3840 x 2160 pixels.

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

  • The recommendations on channel count in typical single-device VDEC scenarios are as follows.
    Table 1 Performance specifications of the Ascend AI Processor on the Atlas 200/300/500 Inference Product

    Resolution

    Total Frame Rate

    Per Channel Frame Rate (n Channels)

    Max Memory Usage per Channel (Using H.265 Input Streams with Nine Reference Frames as an Example)

    4K

    120 FPS

    120 FPS/n (n = 4 is recommended, 30 FPS per channel.)

    About 182 MB

    1080p

    480 FPS

    480 FPS/n (n = 16 is recommended, 30 FPS per channel.)

    About 56 MB

    ≤ 720p

    960 FPS

    960 FPS/n (n = 32 is recommended, 30 FPS per channel.)

    About 35 MB

    The following table uses 1080p input streams as an example to describe the formula for calculating the maximum memory used by each VDEC channel. The memory usage is calculated as follows:
    • Buffer size for the input streams: 6 MB when resolution ≤ 1080p; 9 MB when resolution > 1080p.
    • Frame buffer size for the decoded images: 3 MB when resolution = 1080p. For other resolutions, convert the value proportionally.
    • TMV buffer size for the decoded images: 0.5 MB for H.264 1080p input streams; 1 MB for H.265 1080p input streams. For other resolutions, convert the value proportionally.
    • Number of reference frames: The maximum number of reference frames is used as an example. The system adapts to the actual number.

    Formula

    Scenario

    Memory Usage

    4 MB + Buffer size for the input streams x 2 + (Frame buffer size for the decoded images + TMV buffer size for the decoded images) x (Number of reference frames + 1)

    • Input stream format: H.264
    • Input stream resolution: 1080p
    • Buffer size for the input streams: 6 MB
    • Frame buffer size for the decoded images: 3 MB
    • TMV buffer size for the decoded images: 0.5 MB

    51 MB (nine reference frames)

    26.5 MB (two reference frames)

    • Input stream format: H.265
    • Input stream resolution: 1080p
    • Buffer size for the input streams: 6 MB
    • Frame buffer size for the decoded images: 3 MB
    • TMV buffer size for the decoded images: 1 MB

    56 MB (nine reference frames)

    28 MB (two reference frames)

  • Suggestions on the number of decoding channels and frame rate:

    The specifications in the following table are for reference only. If the number of channels started per process exceeds these listed specifications, memory insufficiency or performance drop may occur, leading to decoding channel creation failures or slow decoding.

    720p indicates a resolution of 1280 x 720 pixels. 1080p indicates a resolution of 1920 x 1080 pixels. 4K indicates a resolution of 3840 x 2160 pixels.

    Typical Resolution

    Suggested Channel Count per Process (Based on the Input Frame Rate)

    -

    Input frame rate ≥ 25 FPS

    20 FPS < Input frame rate < 25 FPS

    15 FPS < Input frame rate ≤ 20 FPS

    10 FPS < Input frame rate ≤ 15 FPS

    Input frame rate ≤ 10 FPS

    ≤ 720p

    32 channels

    32 channels

    32 channels

    32 channels

    32 channels

    1080p

    16 channels

    19 channels

    24 channels

    32 channels

    32 channels

    4K

    4 channels

    4 channels

    6 channels

    8 channels

    12 channels

Performance Specifications (Atlas Training Series Product)

Table 2 Performance specifications of the Ascend AI Processor on the Atlas Training Series Product

Scenario

Total Frame Rate

n-channel 4-process 1080p (n ≤ 32)

n x 30 x 4 FPS

n-channel 4-process 4K (n ≤ 8)

n x 30 x 4 FPS

The following table uses 1080p input streams as an example to describe the formula for calculating the maximum memory used by each VDEC channel. The memory usage is calculated as follows:
  • Buffer size for the input streams: 6 MB when resolution ≤ 1080p; 9 MB when resolution > 1080p.
  • Frame buffer size for the decoded images: 3 MB when resolution = 1080p. For other resolutions, convert the value proportionally.
  • TMV buffer size for the decoded images: 0.5 MB for H.264 1080p input streams; 1 MB for H.265 1080p input streams. For other resolutions, convert the value proportionally.
  • Number of reference frames: The maximum number of reference frames is used as an example. The system adapts to the actual number.

Formula

Scenario

Memory Usage

4 MB + Buffer size for the input streams x 2 + (Frame buffer size for the decoded images + TMV buffer size for the decoded images) x (Number of reference frames + 1)

  • Input stream format: H.264
  • Input stream resolution: 1080p
  • Buffer size for the input streams: 6 MB
  • Frame buffer size for the decoded images: 3 MB
  • TMV buffer size for the decoded images: 0.5 MB

51 MB (nine reference frames)

26.5 MB (two reference frames)

  • Input stream format: H.265
  • Input stream resolution: 1080p
  • Buffer size for the input streams: 6 MB
  • Frame buffer size for the decoded images: 3 MB
  • TMV buffer size for the decoded images: 1 MB

56 MB (nine reference frames)

28 MB (two reference frames)