Computing Power Test

Function

Create a matrix multiplication A(m,k)*B(k,n) and execute it multiple times. Then, calculate the computing power of the AI Core in the entire card, server, or chip, and the real-time power in full computing power mode based on the computation amount and the time used for performing matrix multiplications. For Atlas A2 training product, the computing power of the AI Core in the entire card, server, or chip, and real-time power in full computing power mode are calculated based on computation amount and the time required for performing matrix multiplication and vector multiplication multiple times.

The involved parameters are shown in Table 1and Table 2.

**Table 1** Matrix multiplication parameters
Operator Operation Type	Parameter	Description	Value
FP16 (all device models) BF16 (Atlas A2 training products, Atlas A2 inference products, Atlas A3 training products, Atlas A3 inference products, Ascend 950PR and 950DT products) INT8 (Atlas 200/300/500 inference products)	m	Rows in matrix A	256
	k	Columns in matrix A and rows in matrix B	32
	n	Columns in matrix B	128
INT8 (Atlas 200I/500 A2 inference products, Atlas training products, Atlas A2 training products, Atlas A2 inference products, Atlas A3 training products, Atlas A3 inference products, Ascend 950PR and 950DT products) HF32 (Atlas A2 training products, Atlas A2 inference products, Atlas A3 training products, Atlas A3 inference products, Ascend 950PR and 950DT products) FP8 (Ascend 950PR and 950DT products) MXFP8 (Ascend 950PR and 950DT products) HIF8 (Ascend 950PR and 950DT products)	m	Rows in matrix A	256
	k	Columns in matrix A and rows in matrix B	64
	n	Columns in matrix B	128
FP32 (Atlas A2 training products, Atlas A2 inference products, Atlas A3 training products, Atlas A3 inference products, Ascend 950PR and 950DT products)	m	Rows in matrix A	128
	k	Columns in matrix A and rows in matrix B	32
	n	Columns in matrix B	64
MXFP4 (Ascend 950PR and 950DT products)	m	Rows in matrix A	512
	k	Columns in matrix A and rows in matrix B	64
	n	Columns in matrix B	128

**Table 2** Vector multiplication parameters
Operator Operation Type	Parameter	Description	Value
FP16 (Atlas A2 training products, Atlas A2 inference products, Atlas A3 training products, Atlas A3 inference products)	n	Vector length	32760
FP32 (Atlas A2 training products, Atlas A2 inference products, Atlas A3 training products, Atlas A3 inference products) HF32 (Atlas A2 training products, Atlas A2 inference products, Atlas A3 training products, Atlas A3 inference products) BF16 (Atlas A2 training products, Atlas A2 inference products, Atlas A3 training products, Atlas A3 inference products)	n	Vector length	16380
FP32 (Ascend 950PR and 950DT products)	n	Vector length	16128
FP16 (Ascend 950PR and 950DT products) BF16 (Ascend 950PR and 950DT products)	n	Vector length	32256
INT8 (Ascend 950PR and 950DT products)	n	Vector length	64512

Settings Before the Test

The Ascend AI Processor presets performance thresholds. You are advised to perform the computing power test when the device temperature is stable and lower than 90°C. This prevents frequency reduction from being triggered due to high device temperature, resulting in inaccurate computing power test results.

Parameters

You can run either of the following commands to list the parameters of the computing power test command:

ascend-dmi -f -h

ascend-dmi -f --help

Table 3 lists only a test-specific parameter. For details about other common parameters, see Common Parameters.

**Table 3** Parameter description
Parameter	Description	Mandatory
[-f, --flops]	Measures the computing power of the entire card, chip, or server.	Yes
[-t, --type]	Specifies the operator operation type, which can be fp16, fp32, hf32, bf16, int8, fp8, mxfp8, hif8, or mxfp4. If this parameter is not specified, fp16 is used by default.	No
[--all]	If this parameter is specified, the computing power of the entire server is tested, that is, the sum of the computing power of all NPUs is calculated. This parameter cannot be used with -d.	No
[-et, --et, --execute-times]	Specifies the number of times that matrix multiplication is performed on a single AI Core on a specified chip. If not specified, the default value 60 is used. The value ranges from 10 to 80, with the unit being 100,000.	No
Note: In this document, the input or output device ID is the logic chip ID. You can run the npu-smi info -m command to obtain the logic ID by viewing the value of Chip Logic ID on the GUI. The NPU ID is the physical chip ID.

Example

Perform a computing power test on device 2 by executing a matrix multiplication 6 million times. The default operator operation type is fp16.
ascend-dmi -f -d 2 --et 60
Perform a computing power test on device 2 by executing a matrix multiplication 6 million times. The default operator operation type is int8.
ascend-dmi -f -t int8 -d 2 --et 60

Perform a computing power test on device 3 by executing a matrix multiplication 8 million times.
ascend-dmi -f -d 3 --et 80
Perform a computing power test on Atlas A2 training products with the computing power test type set to hf32.
ascend-dmi -f -t hf32
Perform a computing power test on Atlas A3 training products with the computing power test type set to fp32.
ascend-dmi -f -t fp32 -q
Perform a computing power test on Atlas A3 training products with the computing power test type set to bf16.
ascend-dmi -f -t bf16 -q

Perform a computing power test on Ascend 950PR and 950DT products with the computing power test type set to hif8.

ascend-dmi -f -t hif8 -q

-----------------------------------------------------------------------------------------
    Device        Execute Times        Duration(ms)        TFLOPS@HIF8        Power(W)
-----------------------------------------------------------------------------------------
    0             504,000,000          2320                804.040            377.7
-----------------------------------------------------------------------------------------

Perform a computing power test on Ascend 950PR and 950DT products with the computing power test type set to mxfp4.

ascend-dmi -f -t mxfp4 -q

-----------------------------------------------------------------------------------------
    Device        Execute Times        Duration(ms)        TFLOPS@MXFP4         Power(W)
-----------------------------------------------------------------------------------------
    0             504,000,000          1386                1561.195           373
-----------------------------------------------------------------------------------------

Specify --all to test the computing power of the entire server (FP16 used as an example).

ascend-dmi -f -q --all

If --all is specified, Execute Times, Duration (ms), and Power (W) indicate the average values of other indicators of the entire server. Device is set to all, indicating all NPUs. TFLOPS@FP16 indicates the sum of the computing power of all devices.

------------------------------------------------------------------------
  Device      Execute Times     Duration(ms)    TFLOPS@FP16     Power(W)
------------------------------------------------------------------------
  all         360000000         1702            2509.719      206.625015  
------------------------------------------------------------------------

Table 4 describes the server parameters in the preceding figures.

**Table 4** Parameter description
Parameter	Description
Device	Indicates the device ID.
Execute Times	For the Atlas A2 training products, Atlas A2 inference productsAtlas A3 training products, Atlas A3 inference products and Ascend 950PR and 950DT products, the value of Execute Times is the sum of the number of matrix multiplication times of a single AI Core multiplied by the number of AI Cores and the number of vector multiplication times of a single vector core multiplied by the number of vector cores. For the Atlas training product and other inference products, the value of Execute Times is the number of matrix multiplication times of a single AI Core multiplied by the number of AI Cores.
Duration(ms)	Indicates the time used to complete the matrix multiplication computation.
TFLOPS@FP16	Indicates the calculated computing power. FP16 is the specified operator running type.
Power(W)	Specifies real-time power in full computing power mode. NOTE: You do not need to pay attention to the chip power during the computing power test because the power consumption data is collected periodically and there is an interval between two collections. When the computing power test period is too short, power consumption data fluctuates. Use a more specific power consumption test option to test the power consumption.

To ensure the correctness and accuracy of the test result, perform the computing power test separately.

Parent topic: Performance Test