Hey everyone!
I have a new build arriving soon featuring an RTX 5060 Ti and my primary use case will be running small local LLMs.
My main goal is to lower power consumption and thermals without hurting token generation speeds (t/s). I've read that capping the Power Limit to roughly 150W results in almost no real performance penalty (~1-2%).
My thought was to take this a step further: instead of just sliding the power limit down and letting the GPU throttle clocks unpredictably, I’d like to do a proper manual V/F curve undervolt (similar to a CPU negative offset) to lock in higher core frequencies at lower voltages within that ~150W budget.
I’d love to get your thoughts on two specific things:
Do you agree with this strategy for LLM inference? Is dialing in a manual V/F curve actually worth the effort for memory-bandwidth bound workloads, or does a plain Power Limit cap achieve practically the same result?…