Earlier quoted context omitted.
A lot of the TDP is reserved for running the shader units at full-power. My RTX 3070 Ti only pulls ~110w of it's 320w running CUDA inference on Gemma 26b and E4B.
It's not that it's reserving power, but rather that you hit some bottleneck on a 3070 Ti before running into thermal limits-- it's likely limited by either tensor core saturation or RAM throughput. Running the workload with Nvidia's profiling tools should make the bottleneck obvious.
Generally the bottleneck is RAM throughput. Inference, in particular token generation, especially on a single user instance, is not all that computationally complex; you're doing some fairly simple calculations for each parameter, the time is dominated by just transferring each parameter from RAM to the cores. A 31B dense model like Gemma 4 has to transfer 31B parameters (at 16 bits per parameter for the full model, though on consumer hardware people generally run 4-8 bit quantizations) from RAM to the cores, that's a lot of memory transfer.
Prompt processing or parallel token generation can do a bit more work per memory transfer, as you can use the same weights for a few different calculations in parallel. But even still, memory bandwidth is a huge factor.