Search code examples
pythontensorflowmemorygpuhardware

Incorrect memory bandwidth when using TensorFlow

I was wondering whether this memory bandwidth amount is correct. I have an NVDIA RTX 3090 and, in theory, it's bandwidth should be around 936.2 GB/s. However, when using TensorFlow, it appears significantly lower. Is there any limitation or is this how it should be? My current bus interface is PCIe x16 4.0

enter image description here

Solution

  • 871.81 gibibyte (GiB) ~= 936.1 gigabyte (GB)

    So the bandwidth reported by Tensorflow is about the same (0.1 GB/s difference) as the official specification of the card. I suspect the 0.1 difference is just rounding or precision loss somewhere.