Performance Specifications
This section describes only the reference performance specifications of AI processor for each product. The performance specifications vary according to the memory and CPU configurations of each product. For details about the performance specifications of each product, see the official description on the Ascend community.
This section takes a single device 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, 1080p indicates a resolution of 1920 x 1080 pixels, and 4K indicates a resolution of 3840 x 2160 pixels.
The performance specifications of JPEGD are based on hardware decoding performance. JPEGD hardware decoding does not support the image decoding with three SOS markers. For image formats not supported by hardware, software decoding is used. The software decoding performance for reference is 1080p at 15 FPS. If the output image decoded by JPEGD involves rotation, the performance is lower than the reference values of software decoding. For example, the performance is lower than 15 FPS for 1080p images.
Performance Specifications (
Atlas A2 training product
/
Atlas A2 inference product
)
The following performance data is obtained by invoking the aclrtSynchronizeStream interface once after 10 asynchronous media data processing tasks are delivered on a stream.
|
Input Format |
Input Image Size |
Scenario |
Total Frame Rate |
|---|---|---|---|
|
YUV400/420 |
≤500KB |
n-channel 1080p (1 ≤ n ≤ 28) |
n*512fps |
|
n-channel 1080p (n > 28) |
14336fps |
||
|
YUV440/422/444 |
≤500KB |
n-channel 1080p (1 ≤ n ≤ 28) |
n*256fps |
|
n-channel 1080p (n > 28) |
7168fps |
||
|
YUV400/420 |
≤2MB |
n-channel 4K (1 ≤ n ≤ 28) |
n*128fps |
|
n-channel 4K (n > 28) |
3584fps |
||
|
YUV440/422/444 |
≤2MB |
n-channel 4K (1 ≤ n ≤ 28) |
n*64fps |
|
n-channel 4K (n > 28) |
1792fps |
Performance Specifications (
Atlas 200I/500 A2 inference product
)
The following performance data is obtained by invoking the aclrtSynchronizeStream interface once after 10 asynchronous media data processing tasks are delivered on a stream.
|
Scenario |
Total Frame Rate |
|---|---|
|
n-channel 1080p (1 ≤ n ≤ 2) |
n*256fps |
|
n-channel 1080p (n > 2) |
512fps |
|
n-channel 4K (1 ≤ n ≤ 2) |
n*64fps |
|
n-channel 4K (n > 2) |
128fps |
Performance Specifications (
Atlas inference product
)
The following performance data is obtained when the aclrtSynchronizeStream interface is executed once after 10 asynchronous media data processing tasks are delivered to a stream. The host server is configured as follows: The CPU dominant frequency is Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, and the number of CPU cores is 14.
|
Scenario |
Total Frame Rate |
|---|---|
|
n-channel 1080p (1 ≤ n ≤ 16) |
n*128fps |
|
n-channel 1080p (n > 16) |
2048fps |
|
n-channel 4K (1 ≤ n ≤ 16) |
n*32fps |
|
n-channel 4K (n > 16) |
512fps |
Performance Specifications (
Atlas training product
)
The following performance data is obtained by invoking the aclrtSynchronizeStream interface once after an asynchronous media data processing task is delivered on a stream.
|
Scenario |
Total Frame Rate |
|---|---|
|
n-channel 1080p (1 ≤ n ≤ 16) |
n*128fps |
|
n-channel 1080p (n > 16) |
2048fps |
|
n-channel 4K (1 ≤ n ≤ 16) |
n*32fps |
|
n-channel 4K (n > 16) |
512fps |