Unfortunately this needs to be balanced by transient response of the inductor dictated by current-storage,heat dissipation, and phase interleaved mode as it provides benefits.
This video was already linked and discussed on the previous page.
I'd suspect these use slightly lower memory voltage? It looks like the 8GB/8gbps cards use 1.55V.
Might depend on what rating they are using. The part selection would make more sense if they were running at very high temperatures. FINFET temperature inversion may have led them to design for unusually high thermals. How many other thermally sensitive parts are rated beyond spec? The card design makes more sense if you think about a blower pushing air across a GPU at 100C or more before hitting the VRMs.
You can't underclock the memory; WattMan doesn't allow it, and 3rd party tools don't yet speak the new power API. And as someone else has already mentioned, there's also the power/voltage issue. This is why AMD gave reviewers the 4GB/7Gbps BIOS for their reference cards, as it was the only proper way to get the memory down to 7Gbps.
You'd be surprised what happens to a loaded-down ROP partition. Filling the ALUs won't always get you max power consumption.
Going off these numbers the 480 is always attempting to maintain 80C in both Crisis 3 and FurMark. Based on that paper I linked above the card should start out running hot and slow until it reached 80C. Which it appears to be doing rather well. Once that point is reached in theory it would start reducing the voltage while possibly increasing frequencies. That part doesn't appear to be happening reliably. If it did adjust voltages, the whole PCIE issue would likely go away because it would be backing off that power wall. A 40% reduction in voltage with the same frequencies would drastically lower power by nearly two thirds and I doubt frequencies could increase enough to compensate for that.
80C is when the fan ramps up. Until you reach the fan maximum (i.e. how loud you're going to allow it to be), it's not going to back off on frequency due to temperature. I never reached the fan speed maximum in my testing of games, though AMD doesn't have NV's neat API flag that tells you why the card is throttling.
Not sure how it handles the voltage, but the power draw is actually measured by the voltage drop over a shunt resistor in the 12V net. So the card can measure the actual board power rather reliably, rather than going by estimations/heuristics.
Can you quantify how much of that difference is due to possibly lower framerates for the 4-gbds-BIOS'ed card and thus lower system-as-a-whole load?
To add to that: For maximum power consumption, you need to take into account a couple of things. The different parts of the chip from top to bottom in the pipeline should be doing at least some work to keep them from powering down. While pure compute loads as in Sisoft Sandra do tend to punish the ALUs quite good as well as the internal memory system, actually setting up stuff, drawing a part of it and computing things at the same time often does drive up power harder. There's a fine balance between the higher load on extreme 4K settings and achieving a high Framerate (ROP, Geometry throughput) as well.
The framerate only differs by ~2% at 1080p, so I'd say a couple of watts on the CPU side, and another couple of watts on the GPU side. FurMark may be more relevant, since it's a cleaner load.
