I have heard that some AI GPU boards have a function that automatically optimizes GPU (neural network) calculations with hardware features and stores parameters in internal non-volatile memory.
my questions are:
Does Jetson AGX Orin have such a function?
Could this be one of the reasons for the GPU temperature increase?
If so, is there a way to initialize the parameters stored in the non-volatile memory?
I previously posted a question about rising GPU temperatures here .
I am still investigating.
AGX Orin GPUs now exceed 99°C and start Thermal Throttling.
On Jetson, we recommended TensorRT for neural network optimization.
Usually, optimization reduces some non-necessary bottlenecks (ex, memory copy) to allow high GPU occupancy.
Since GPU in a much higher loading, the power consumption is expected to increase.
I am testing the cooling performance of the device, using the stress command for the CPU and gpu-burn for the GPU, always keeping it at 100% load. Therefore, I am not interested in optimizing, but rather in maintaining a stable maximum load.
I am currently experiencing problems with the GPU temperature rising. I have not changed any of the test conditions or software settings, so I wondered if there was an unintended optimization function in the hardware, and so I asked the question.
Are there any implicit hardware optimizations like the ones I mentioned in my post?
