Radeon Software 16.7.1 Adjustments

AMD released a new driver today that they hope fixes the issues with power consumption. Does it?

Last week we posted a story that looked at a problem found with the new AMD Radeon RX 480 graphics card’s power consumption. The short version of the issue was that AMD’s new Polaris 10-based reference card was drawing more power than its stated 150 watt TDP and that it was drawing more power through the motherboard PCI Express slot that the connection was rated for. And sometimes that added power draw was significant, both at stock settings and overclocked. Seeing current draw over a connection rated at just 5.5A peaking over 7A at stock settings raised an alarm (validly) and our initial report detailed the problem very specifically.

AMD responded initially that “everything was fine here” but the company eventually saw the writing on the wall and started to work on potential solutions. The Radeon RX 480 is a very important product for the future of Radeon graphics and this was a launch that needs to be as perfect as it can be. Though the risk to users’ hardware with the higher than expected current draw is muted somewhat by motherboard-based over-current protection, it’s crazy to think that AMD actually believed that was the ideal scenario. Depending on the “circuit breaker” in any system to save you when standards exists for exactly that purpose is nuts.

Today AMD has released a new driver, version 16.7.1, that actually introduces a pair of fixes for the problem. One of them is hard coded into the software and adjusts power draw from the different +12V sources (PCI Express slot and 6-pin connector) while the other is an optional flag in the software that is disabled by default.

Reconfiguring the power phase controller

The Radeon RX 480 uses a very common power controller (IR3567B) on its PCB to cycle through the 6 power phases providing electricity to the GPU itself. Allyn did some simple multimeter trace work to tell us which phases were connected to which sources and the result is seen below.

The power controller is responsible for pacing the power coming in from the PCI Express slot and the 6-pin power connection to the GPU, in phases. Phases 1-3 come in from the power supply via the 6-pin connection, while phases 4-6 source power from the motherboard directly. At launch, the RX 480 drew nearly identical amounts of power from both the PEG slot and the 6-pin connection, essentially giving each of the 6 phases at work equal time.

That might seem okay, but it’s far from the standard of what we have seen in the past. In no other case have we measured a graphics card drawing equal power from the PEG slot as from an external power connector on the card. (Obviously for cards without external power connections, that’s a different discussion.) In general, with other AMD and NVIDIA based graphics cards, the motherboard slot would provide no more than 50-60 watts of power, while any above that would come from the 6/8-pin connections on the card. In many cases I saw that power draw through the PEG slot was as low as 20-30 watts if the external power connections provided a lot of overage for the target TDP of the product.

Starting with the 16.7.1 driver, AMD will automatically reprogram the power controller on the RX 480 to better divide the power draw from the 6 available phases on the reference cards. This is a process that occurs at each and every boot, it is not a permanent change to the VBIOS. It’s possible, and likely, that future cards and partner cards may have this change integrated at a lower level, negating the need for the driver to recognize and update the controller logic. But for now, with launch RX 480s in the wild, that’s how the process works.

As I understand, what AMD is doing now is very similar to what The Stilt in the Overclock.net forums attempted earlier in the week. Power phases 1-3 that source +12V from the 6-pin connection are now given more time than phases 4-6, thus shifting the weight of power draw towards the 6-pin connector. We’ll be able to calculate that exact ratio when we show you the power consumption data from our testing, but the goal is draw less power from the PCI Express slot and more over the 6-pin connector while maintaining the exact same power and performance profiles. To be clear: this fix will not affect performance and my testing shows that to be the case.

You have to wonder why this wasn’t the direction taken by AMD engineers initially. As you will find in our results, the 6-pin connection is definitely drawing more than the 75 watts that it is rated at, but the 6-pin cabling and connectors are actually rated at 8-9A per pin, rather than 5.5A total for the PEG slot. If you are going to draw more power than rated over one of the two options, it’s clear that the over engineered and directly fed 6-pin connection is the way to go. I can only assume it was an oversight on the board team that allowed it to happen. (Also, let’s not forget that using an 8-pin connection and weighting it towards that would have prevented both issues.)

Compatibility Mode

Also included with the 16.7.1 driver is a new toggle in the global settings called Compatibility Mode. It’s an absolutely mind-numbing name for a feature that simply does one thing: lowers the total target power draw for the GPU.

This fix was actually the first we tested from AMD, though it wasn’t enough to alleviate the problems with power draw from the PCI Express connection. This adjustment does nothing to change the weighting of power draw from the two +12V sources and instead only focuses on lower total power draw of the GPU. Yes, this does mean that there are going to be some cases where performance drops, though I have seen articles talking about performance increases with undervolting.

AMD indicates that we should think of this setting as a secondary solution, one that is NOT enabled by default, for any users that might be overly concerned about power and current draw on their motherboard or power supply. As I will show you in our testing, the differences in power draw are definitely measurable, and performance seems to be impacted minimally in the couple of spot checks we have done. 

« PreviousNext »