NVIDIA GeForce RTX 3090 Founders Edition Review: Between Value and Decadence - When Price is Not Everything | Page 16

Power consumption and compliance with standards

When it comes to power consumption, there are actually no big secrets, because what NVIDIA specifies as TGP is largely complied with. With about 16 watts in idle, one is in the range of many board partner cards of the GeForce RTX 2070 Super to RTX 2080 Ti, even if relatively high. The partial load ranges are realized relatively economically, which could also be due to the better controlled clock rate. However, the card reaches 360 watts under real full load in the Witcher 3 at Ultra-HD and full warming, i.e. 10 watts more than predicted and stated in the specs.

But this is not the whole truth, because the power consumption is very resolution-dependent. Why is that? Ask the CPU, which in some situations already slows down the fast sweeper a little bit. Not in every game, but in some it happens. We had already analyzed which exactly these are when we were talking about the combined power consumption with the CPU and its individual listings. But if you add up the power consumption in Ultra-HD, which I measured in all games over the entire runtime, then everything is back at the 350 watts given for the TGP, a spot landing, because the CPU doesn’t limit any more. First of all, the average over all games and graphics cards:

The following table now shows the maximum values of the respective load scenarios and the averaged gaming glass. The value for Gaming UHD is the highest value measured in Witcher 3 because the game is still the spearhead of games in terms of power consumption.

Which brings us to the specifications, because at the end of the day the firmware clearly regulates what is possible and what is allowed to be done at maximum and minimum. A maximum of 400 is allowed for the VU, which can be retrieved with the appropriate software, but is also somewhat pointless.

Let us now compare the Voltage NVVDD with MSVDD. which can be between 0.7 and 1.2 volts at ampere (technically speaking) and which for safety reasons rarely exceed the 1-volt limit with normal consumer cards. You can see very clearly from the curve that the slightly warmer card operates with less voltage changes.

Let’s now move on to the load on the motherboard slot, whose load is specified by the PCI SIG as 5.5 amps. This results in a maximum power of 66 watts at 12 volts. You can see very clearly that under full load exactly 5.5 amps are reached, which is of course still permissible, but already borderline. However, this is slightly exceeded with 5.6 to 5.7 A if the maximum power limit is used. This is not very nice and it is probably also due to the shifted balancing of the added 10th place.Phase that would probably have been better served from the 12-pin.

The somewhat more detailed curve for gaming and torture looks like this, whereby the intervals with the long 20 ms are still very granular.

Transients and power supply recommendation

As I have already proved in detail in my basic article “The battle of graphics card against power supply – power consumption and load peaks demystified”, there are also higher loads in the millisecond range, which can lead to unexplainable shutdowns in the case of unfavorably designed or improperly equipped power supplies. The TBP (Typical Board Power) measured by the graphics card manufacturer or the reviewers alone does not really help in this case for a stable system design.

Peaks with intervals between 1 and 10 ms can lead to shutdowns in very fast reacting protective circuits (OPP, OCP), especially in multi-rail power supplies, although the average power consumption is still within the standard. For this card, I would therefore calculate at least 460 to 500 watts as a proportion of the total secondary power consumption of the system with a normal OC with up to 400 watts as the power limit as the graphics card load, in order to have enough reserves for the worst case scenario. A short excerpt with higher resolution shows us now the 20 ms measurements (10 μS intervals), as I let them run automatically for value determination: