Correct. The main bottleneck with LLM inference is, and have always been, memory bandwidth.

TPS = active weights in GB / your memory bandwidth.

That’s it for decode. That’s all.