Here are my findings after spending today testing undervolting my Sapphire Vega 64. First, a few things I learnt while fiddling with the settings:
- Core voltage settings in WattMan will only take effect if they are equal to or above the memory voltage, so you should set memory voltage to no more than the lowest core voltage. This doesn't affect the actual memory voltage at all, so it is presumably a bug. Fortunately I had read about this before, otherwise I may not have worked it out myself.
- Using the render test in GPU-Z as an easy way to kick the GPU into the highest power state, I could see that reducing core voltage from the default "Custom" 1200mV initially increases the actual core speed. As it is reduced further, the core speed starts to fall. This doesn't seem to be affected by the current temparature, or increasing the power limit, so I'm not sure what's limiting it.
To obtain the following numbers, I used the game
Elex, as it happens to be what I am playing at the moment. After changing settings in WattMan, I left the game paused for a few minutes to let the temperature stabilize, then used HWiNFO64 to record the averages while I walked along the main road from
Sandy Pines to
Old Windfarm. The AC power figures are from my basic plug-in power meter. It can't average over time, so I just recorded a single figure when standing still at a particular spot.
Code:
State 6/mV State 7/mV Power limit/% GPU Clock/MHz VDDC/V Hot Spot/°C Fan Speed/RPM Chip Power/W AC Power/W
[Balanced] [Balanced] [Balanced] 1561 1.063 87 1691 261 392
1000 1100 0 1553 1.047 83 1564 244 370
1000 1080 0 1552 1.030 81 1497 233 360
1000 1080 50 1552 1.029 80 1554 233 359
I had a display driver crash when trying 1070mV for state 7, so the lowest setting I recorded was 1080mV. I won't know if that's completely stable until I have given it a few weeks of stress. It's hard to know what the optimum for state 6 is without a lot more benchmarking, but locking to state 6 as the maximum shows that going below 1000mV causes the actual core clock to go down fairly fast. I recorded an extra set of numbers with the power limit raised to demonstrate that power is not limiting core speed.
There is a nice reduction in chip power of nearly 30W as core voltage is reduced, which matches the measured change in AC power, allowing for PSU inefficiencies. As you would expect to follow on from that, both temperature and fan speed are reduced too. There is a very slight reduction in average clock speed, but small enough that it's almost lost in the noise.
In conclusion, it's well worth undervolting, but working out the best settings is difficult because they appear to be treated as suggestions rather than hard values.