TDA4VM: Inference benchmarking statistics - DDR BW Per Image

Hi,

Responses will be delayed until next week due to the US Thanksgiving holidays.

regards

Suman

Hi,

I still have this problem. Do you have any idea where it might come from?

Could it be related to an issue that prevents me from making inferences via TIDL-RT on the C7x core?

Thanks,

Azer

Hi Azer;

I don't think we should compare with your co-worker's two-years-ago results; unless you still have the exact software and setup to compare.

By looking at your screen shot, DDR usage was un-real high. What is the software you used to get this screen shot? 

I will forward your question to the DDR expert. We will look at the TIDL-RT issue later.

Thanks and regards

Wen Li 

Hi,

 I get the same result on EVM and on the cloud, so I will detail how I achieve this result via the cloud for now.

Here is the model I am using: /home/root/edgeai-modelzoo/models/vision/classification/imagenet1k/mlperf/mobilenet_v1_1.0_224.tflite.

I launch the modified vcls-tfl notebook to run on ARM only.

And then using the following code, I get the benchmark results : 

from scripts.utils import plot_TI_performance_data, plot_TI_DDRBW_data, get_benchmark_output, print_soc_info
stats = interpreter.get_TI_benchmark_data()
fig, ax = plt.subplots(nrows=1, ncols=1, figsize=(10,5))
plot_TI_performance_data(stats, axis=ax)
plt.show()

tt, st, rb, wb = get_benchmark_output(stats)
print_soc_info()

print(f'{selected_model_id.label} :')
print(f' Inferences Per Second    : {1000.0/tt :7.2f} fps')
print(f' Inference Time Per Image : {tt :7.2f} ms')
print(f' DDR usage Per Image      : {rb+ wb : 7.2f} MB')

I'm doing exactly the same on EVM, with other models too.

Regards and thanks,

Flora

Ok, thanks for providing the details. I will try your script. I may need some time to setup. 

Regards

Wen Li

Hi, 

I noticed I only have this anomaly when running on ARM only. DDR results are normal on C7x+MMA.

Regards,

Azer