Why do AMD’s actually measured values at the pcie connectors deviate so extremely from the results read out via software and why is the error rate in short-term measurements on NVIDIA graphics cards still so high? Where do the values read out by the tools actually come from, and what is the fundamental difference between the two graphics card manufacturers in the end? Today’s article is intended to show you where the respective pitfalls lie with the free tools and what is really sufficient for accurate estimation of power consumption (and what is not).
Because there are really huge differences between AMD and NVIDIA, which I would like to explain to you as understandably as possible based on the board design and the approach of both manufacturers. In addition, I will give you some hints on how you can or must supplement or extrapolate the data read out by the tool. Because already at this point we have to separate and distinguish properly between real measurement and simple readout via software interfaces. That’s because those are actually not comparable, at least with the Radeons. Therefore, I will briefly start with the measurements again (for a better understanding of the task), since the necessary basics have to be talked about first.
Why a measurement is irreplaceable and yet not everyone needs it every day
For the normal user, the measurements, as I make them in the laboratory, are important as a pure source of information and also an essential basis for understanding. Above all, the data on the load peaks, the load change behavior and the estimation of the appropriate dimensioning of the power supply are findings that can only be generated in a suitable laboratory with the necessary technical and financial effort. As a customer, however, you don’t have to make this effort at all, because you already get everything delivered free of charge by third parties (like me).
You can measure it (like me) with two coupled oscillographs, current clamps and probes for the respective voltages (up to 8 channels in total) to measure, record and evaluate even the finest load peak at intervals far below one millisecond. Of course, this allows deep insights into the behavior of the graphics cards at different loads, but is almost useless for individuals without a professional evaluation and summary due to the sheer amount of data. I, too, first have to simplify everything and also prepare it graphically in such a way that you can understand anything at all. For various reasons, I won’t go into all the details, not even the current setup in the lab, because I too have one or two trade secrets. But at the end I have a collection of links for you, where I describe such things in more detail. Here that would be too confusing.
What must be determined for all measurements, however, are the currents flowing on the 12-volt lines and the corresponding voltages at each of the terminals (since these values also fluctuate extremely). However, the smaller the measurement intervals, the more you have to adapt technically. This is because the current clamps also have built-in electronics, which leads to a certain time delay, while the voltages are, after all, recorded directly at the input of the oscilloscope. So here a complex calibration and the entry of these delays in the firmware menus for the voltage inputs are mandatory.
There is nothing worse than when time-shifted values are then multiplied together (voltage times current)! This is because a current peak does not always fall on a high or at least average voltage. What is applied to the 12V rails is anything but a clean DC voltage! So you can also measure a lot of mischief (up to the dreaded aliasing) if you don’t pay close attention to what you are actually doing.
Of course, it is also possible to measure all this using a very elaborate MCU-based shunt measurement setup for all current-carrying lines, but this loses some of the resolution. There are currently hardly any solutions that can still measure below one millisecond, since everything is usually connected via USB. The polling rate already limits the flow of information considerably. The advantage, however, is that you don’t have to make a medium-high, five-figure investment, but can be happy with up to 1000 euros, depending on your requirements and effort.
ElmorsLAB’s PMD and NVIDIA’s PDAT also fall into this cateory of cheaper shunt solutions. The resolution is not particularly high, but for the daily recording of the relevant power consumption and a good momentary checks, these hardware solutions are definitely enough. Those who want to invest some money here will certainly not be disappointed. And since such solutions monitor the incoming 12V lines externally, everything works equally well for AMD and NVIDIA graphics cards, as does my Oscilloscope solution.
Currents also flow on the 3.3-volt rail!
However, the 3.3 volt line is still used by some cards (with up to 5 watts and more!), but it isn’t detected by any software tool and even by NVIDIA’s PCAT board, which can mean a coarser measurement error (and also software readout error) of a few watts. Reading the 3.3-volt rail via sensor is generally impossible, as we will see in a moment on the following pages. However, I can already spoil that NVIDIA’s board partners like to use this rail to take various low voltages out of the overall 12V line budget. This is nothing more than a cheat against NVIDIA’s control rage, but I’ll get to that in a moment.
The picture above shows the evaluation of the data for a rather smaller, but current NVIDIA graphics card, where sometimes 4.5 watts (and more) flow over the 3.3 volt connection and are not included in the total board power (TPB) of the graphics card as the to be respected power limit. This also explains why, for example, two cards from different manufacturers can clock at different speeds despite identical power consumption on 12 volts and similarly high-quality components. Especially with smaller graphics cards up to 150 watts TBP, something like this is definitely not only measurable.
657 Antworten
Kommentar
Lade neue Kommentare
Urgestein
1
Urgestein
Veteran
Mitglied
Veteran
Urgestein
Urgestein
Urgestein
1
Urgestein
1
Urgestein
Urgestein
Urgestein
Urgestein
Veteran
Urgestein
Urgestein
Alle Kommentare lesen unter igor´sLAB Community →