Load on DirectML and TensorRT

[animeojisan]

Thank you for the wonderful app!
Is it by design that MPV-Lazy has a lower load with DirectML compared to TensorRT?

The same result occurs when comparing SR_ESRGAN_DML.vpy and SR_ESRGAN_NV.vpy. The load is also significantly lighter with DirectML in ETC_UAI_DML.vpy versus ETC_UAI_NV_TRT.vpy.

Generally, I would think TensorRT would process faster, wouldn't it?

[hooke007]

I don't know Microsoft or nvidia does what in their API.
It is true trt is always faster and have better, stable image quality in my tests.

[animeojisan]

For example, when using the default settings of SR_ESRGAN_DML.vpy and SR_ESRGAN_NV.vpy on a YouTube anime, the GPU usage rate displayed in Task Manager hovers around 20% for DML and about 25% for NV. Is this normal behavior?

Does this mean that the GPU usage rate in Task Manager is unrelated to processing speed?

[hooke007]

the GPU usage rate displayed in Task Manager

It was never accurate.
It may be no sense because they use different hardware cores to implement the features. Also, if you switch to vulkan gpu API, the gpu usage could be different. But it doesn't mean which API could really save power.

[hooke007]

You can try gpu-z to observe the GPU consumption.

[animeojisan]

Thank you for your response! Indeed, upon observing with GPU-Z, DML is hovering around 23%, and NV is at 15%. The results are satisfactory and make sense to me.