The walkthrough configures torch.profiler with ProfilerActivity.PrivateUse1, NeuronConfig, and a Neuron trace export callback, then inspects the resulting data in Neuron Explorer.
The example profiles three implementations of RMSNorm using the same inputs and profiling schedule. A naive version made from individual PyTorch operations recorded 2.355 ms of Neuron device time across six operator calls. torch.rms_norm lowered to one fused Neuron kernel and recorded 403.3 microseconds. A naive function compiled with torch.compile(backend="neuron") ran as one compiled region at 375.5 microseconds.
The post also describes trace artifacts: a compiled Neuron executable is stored in a NEFF file, while its hardware instruction trace is stored in a paired NTFF file. Neuron Explorer can display framework dispatch, runtime scheduling, and NeuronCore execution.
The source does not establish performance for complete models, other operators, or other Trainium configurations. Its figures are measurements from this RMSNorm example, not a general Trainium benchmark.