Underclocking the AMD Radeon RX 6950 XT

In May 2023 I built a gaming rig, the first PC for myself (and still own) for over a decade. Extortionate prices were falling after the crypto boom of lockdown and particularly GPU scalping. I'd decided on AMD as their CPUs reigned for gaming prowess and their GPUs offered better price/performance. Early one week a PowerColor Radeon RX 6950 XT was reduced from around £1000 to £620 so decided to take the plunge. Having previously settled on the 5800X3D, 32Gb RAM and either an RX 6800 or RX 6800 XT due to their performance and efficiency, with no desire to run a kilowatt PSU - I pay the elecricity bill(!) - have the initial outlay or associated heat and noise of the components, the RX 6950 XT was therefore marginally over specification for my Corsair RM850x. So, unconcerned about squeezing every frame per second out of the system, decided to underclock it.

I'd heard about undervolting AMD products but never tried it myself. Normally overclocking Intel CPUs requires upping the voltage to keep it stable at whatever speed you're aiming for, there are subtleties of course and you can use offsets (remembering VDroop on my i5 Sandy bridge). Understanding that you receive diminishing returns is important, because the further you push hardware toward it's limit heat output and power draw increase at a higher rate for each step in frequency. Also keep in mind you're only tuning one component of an entire system and this will increase overall performance of the machine less than the percentage step you're applying to it due to bottlenecks elsewhere.

Having viewed my fair share of tuning guides on exactly what BIOS settings to tweak (in this case driver options), I've never seen a good one on how to decide how much to change those settings or why.

Firstly installed 3DMark which has a perfectly viable demo version on Steam. Whilst I don't personally rate artificial benchmarks over in-game statistics, it did provide comparable figures to evaluate system performance. Plus used by the world leading overclocking enthusiasts to compare systems so can't be all bad.
Then HWMonitor from CPUID to measure power draw, and started running some benchmarks. As with all hardware tuning, the first thing is to run one stock - this is important as it gives you a baseline for comparison.
Using AMD Catalyst settings, first change I attempted was to drop the GPU voltage without changing other settings. This did not go well and the card refused to maintain stock speeds with any significant undervolt whatsoever i.e. crashing during the benchmark.
Next step was to start reducing frequency. Attempting to drop voltage along with clock rate resulted in the card responding poorly. I'd lost the silicon lottery here, but it didn't overly matter as not aiming for maximum performance - rather a decent drop in power draw without affecting framerate too much. Having just built the system after all so no desire to leave a shedload of performance on the table, but thinking this card (being a top of the range model) would be pushed beyond it's "sweet spot"; into diminishing returns territory. So was going to pull it out from there.
Continuing to drop the clock speed, noted 3DMark score each run alongside average power draw. Min/max frequencies set 100MHz apart as didn't want any auto downclocking to affect the results, retaining this range afterward as apparently helps reduce microstuttering.

Result

You can see from the graph there seems to be a fairly clear curve. Leftmost point it's most efficient - 86% system performance for 54% (195w) GPU power. I appreciate this isn't overall system consumption, but as the GPU is easily the most power hungry device in the machine this is significant. Far right it's at stock, max 3DMark score and pulling about 360 watts.

What I've then done - as much to check my own sanity as anything - is plot a tangent on the curve. The point of tangency is indeed located where my final settings occur; I settled on 96% system performance at GPU draw of 268w - a 92W saving and 74% of the original factory setting.

More importantly though, is the heat and noise I've reduced from the component - not to mention increasing it's lifespan.