Performance bottleneck in Sionna channel generation (CPU-bound?)

[natanzi]

Hi,
I’m running Sionna 1.2.1 and I’m seeing very low GPU utilization during channel generation (CDL/TDL, 38.901).

Before I debug further, I want to confirm whether my understanding is correct:

• Are channel models in Sionna (CDL/TDL/TR 38.901) fully executed as TensorFlow ops on the GPU, or do some parts still fall back to CPU kernels? My observation is that the channel generation becomes a CPU bottleneck even when a GPU is available.

• Does Sionna currently support multi-GPU acceleration for channel generation and link-level simulation?
  From the documentation it seems that only single-GPU execution is supported, but I want to confirm if tf.distribute.MirroredStrategy or other TensorFlow multi-GPU strategies are usable with Sionna.

Is there any recommended approach to avoid the CPU bottleneck (e.g., switching channel generation to pure TF GPU ops, using graph mode, larger batch sizes, etc.)?

Any clarification or recommended optimization strategy would be very helpful.

Note: I have access to H100 GPU.

Thanks!

[merlinND]

Hello @natanzi,

Please share a minimal reproducing script and the required input data to reproduce the issue your are observing.