search

ARM-software / armnn

Arm NN ML Software. The code here is a read-only mirror of https://review.mlplatform.org/admin/repos/ml/armnn
https://developer.arm.com/products/processors/machine-learning/arm-nn
MIT License
1.14k stars 307 forks source link

The mali gpu computation capability is only used no more than 15% when inference resnet50 with ArmNN #760

Closed zxros10 closed 2 months ago

zxros10 commented 4 months ago

I use the application mali_cl_peak_flops_example to test my mali gpu computation capability. Then I test the resnet50 inference with the gpu. I monitor the gpu utilization rate with Streamline, is nearly 100% (GPU Activate). The resnet50 computation(MACs) can get by onnx_tool. Then I compute the gpu capability utilization rate:

resnet50_model_computation / inference_time_after_warm_up / mali_gpu_computation_capability x 100%

The result is less than 15%, I think is too low. The resnet50 is 1x3x224x224, and download from net, not a special one. Do you test this rate before ?

morgolock commented 3 months ago

Hi @zxros10

What format is the resnet50 model you are running? Could you please provide more details about how you run the model? Do you use ExecuteNetwork?