Problems with power consumption in idle mode and ASPM
Active State Power Management (ASPM) is a mechanism of the PCI Express standard that was developed to reduce the power consumption of PCIe devices when idle. It enables the connection between the system and the connected device to be set to energy-saving states without impairing the basic functionality. There are different energy-saving levels that vary depending on the inactivity of the connection. However, the Intel Arc B580 is heavily dependent on the configuration and the supported technologies of the overall system in terms of power consumption. Effective use of ASPM can make a considerable difference here, as the power consumption in idle state is significantly influenced by the activity of the PCIe connection. However, optimizing idle power consumption requires interaction between hardware, firmware and software. The activation of ASPM in the BIOS, the use of up-to-date drivers and a matching operating system make a decisive contribution to operating the GPU more efficiently. Such a configuration makes the Intel Arc B580 not only powerful, but also significantly more energy-efficient in operation.
Measured values
Without ASPM, the PCIe connection remains continuously active, which can lead to relatively high power consumption in idle mode. By activating ASPM, however, this consumption can be significantly reduced, as the system can set the GPU to lower energy-saving modes. In practice, this means that my power consumption can drop from well over 30 watts without ASPM to around 12 watts with ASPM, provided the mainboard and drivers are configured correctly. The whole thing also depends on the resolution and refresh rate and with Full HD and 60 Hz you will certainly end up well below 10 watts.
It is interesting that the specified maximum value is exceeded by almost 8 watts despite monitoring. However, the monitoring on the card only measures the 12-volt rails, while I also measure the 3.3-volt rail, which is never without background noise. But it’s fair to say that Intel complies with the specifications. However, I think the partial load values in particular are still too high. But that’s more to do with the architecture, there’s certainly room for improvement in the next iteration.
Load distribution between the PCIe slot and the PCIe sockets
The distribution of the power supply shows that the card has hardly any reserves in terms of TBP and the performance of the rails. While the 8-pin connector is not fully utilized, the mainboard slot is quite close to the limit of what is permitted with up to 5.1 amps. You can leave it that way.
All in all, Intel has solved the load distribution quite well here and I have the link to the basic article on AMD’s current telemetry for a comparison for anyone who is curious. Intel works more like NVIDIA:
Load peaks and capping
Let’s first take a look at the flowing currents. Measurements were taken at rough 20 ms intervals, i.e. around 50 times per second, in order to simulate the load on the supervisor chip of the power supply units (switch-off). We can see that ALL load peaks are capped at 14 A at the latest, which is still OK.
Nevertheless, we still need to take a look at the voltages, or the product of voltage and current flow. I already wrote that I measured at three different power supply connections, even though all three connections meet again somehow on the graphics card board. What we can now see here as much clearer fluctuations and peaks is due to the power supply unit overvolting a little in places and therefore to the voltage and not the currents. This is due to technical reasons, but is not a problem. However, we can also see that the few peaks of up to around 250 watts at this measurement resolution are not caused by the flowing current (graphics card), but actually result from the power supply!
The Torture test is hardly any different, even if you can see the lower peak values and, above all, the drops. The average, on the other hand, actually rises slightly.
If we now add the voltage to the equation, we see a stronger ripple, which in turn results from the somewhat jittery operating voltage. However, to the power supply’s credit, it must also be said that this affects all current products from all manufacturers and can hardly be avoided.
But because I still want to know exactly, I resolve the whole thing even higher and take 20 ms as the total runtime. The intervals of 10 microseconds can still be measured reasonably and we can also see the voltage here as a green curve, the average value of which is just over 12 volts, but which nevertheless still alternates slightly within the permissible range.
If you then convert this to the power consumption in watts, you get this picture:
I did the same thing again for the Torture Loop, where we can admire the regular drops. First of all, the currents again, but with lots of strange, sporadically recurring drops on every climb. This looks like a violent hiccup before the power is really throttled back shortly afterwards.
And then the total wattage again:
Power supply recommendation
Now we come to the point that makes the expected sensation of exploding power supplies completely absurd. Even IF the card is hopelessly overpowered, nobody actually needs ATX 3.0 power supplies over 1000 watts for such cards, unless the CPU consumes more than 400 watts. This is really a pure job creation measure for the starving power supply industry and only satisfies the sick imagination of some standardization fetishists. You really have to put it that harshly. So you should always stay below 450 to 500 watts, even together with the CPU, if you count up to 10 ms. Because that’s what the power supply units still “see”
My power supply recommendation for the Intel Arc B580 12GB is that a modern 550-watt gold or platinum power supply should get you there quite safely. If you want to overclock, you should plan for 50 watts more, which is especially true for the board partner cards.
- 1 - Introduction and important foreword
- 2 - Teardown: cooler design, material analysis and fan in detail
- 3 - Teardown: circuit board, topology and components
- 4 - Thermal paste from the hell
- 5 - Pad or putty? That will be interesting!
- 6 - Paste vs. pad and putty, high-res thermography, noise
- 7 - Power consumption and load peaks
- 8 - Summary and conclusion
43 Antworten
Kommentar
Lade neue Kommentare
Veteran
1
Veteran
Urgestein
Urgestein
1
Veteran
Urgestein
Veteran
1
Veteran
Neuling
Veteran
1
Urgestein
Urgestein
1
Mitglied
Veteran
Alle Kommentare lesen unter igor´sLAB Community →