Undervolting is the closest thing to a free lunch in this hobby. You feed the card less voltage, it makes less heat, it pulls fewer watts, and on a lot of memory-bound mining workloads your output barely drops or even climbs because the card stops throttling. Lower bill, cooler room, longer-lived hardware. The catch is that it takes an afternoon of patient testing, and most people skip that afternoon.
The idea is simple. Cards ship with more voltage than they need, because the factory has to guarantee stability across millions of chips, including the weak ones. Your specific card is almost certainly better than worst-case, so there is headroom to pull voltage down. Less voltage means less power, and power is what becomes heat.
The stock-clock method
On Nvidia cards the cleanest approach uses MSI Afterburner’s voltage-frequency curve editor. Open the curve with Ctrl+F. It looks like a scatter plot of clock speed against voltage. What you want is to lock the core to a modest clock at a low voltage, so the card stops chasing high boost clocks it does not need.
Pick a point on the curve, say 800mV, and drag it up to a core clock around 1,300MHz. Then flatten everything to the right of that point so the card never runs past that voltage. Apply it. The core now runs at roughly 1,300MHz while sipping 800mV instead of the 1,050mV it wanted at stock.
Then push the memory. Mining is memory-bound on most algorithms, so memory clock is where your output comes from. Raise the memory clock offset in steps of 50MHz, testing stability at each step, until you find the wall, then back off 100MHz for a safety margin. Watch memory junction temperature the whole time, because faster memory runs hotter even as the core cools.
Real numbers
Take an RTX 3070. At stock it might pull about 220W at the wall for the card and run the core hot. Undervolted with a locked core around 1,100 to 1,300MHz and a memory offset near +1,100MHz, that same card commonly lands near 130W while holding full mining output, sometimes a touch higher because it stopped throttling. That is roughly 90W saved per card.
On a six-card rig, 90W each is 540W off your wall draw. Run that 24/7 for a month and you have cut around 388 kWh. At $0.20/kWh that is about $78 a month, every month, for one afternoon of setup. The cards also run 10 to 20 degrees cooler, which means the fans spin slower, wear less, and the whole rig gets quieter.
Testing and stability
Change one thing at a time. Drop the core, confirm output holds and the card is stable for an hour. Then raise memory in small steps. If the driver crashes, the screen artifacts, or your miner throws invalid or rejected shares, you have gone too far on memory. Back off 50 to 100MHz.
Invalid shares are the tell that matters. A card can look stable and still be quietly producing garbage that gets rejected, which is worse than a clean crash because you keep paying for power while producing nothing usable. Let a new setting run for a few hours and watch the rejected-share count before you call it done.
Save your profile once it is dialed in, and set it to load on startup, because a driver update or a reboot will wipe it and the card will quietly go back to pulling full power. Check your temps and wattage again after any driver update. Ten minutes of verification protects the afternoon you spent tuning.