Wednesday, July 8th 2015

AMD Releases PCI-Express Power Draw Fix, We Tested, Confirmed, Works

Earlier today, AMD has posted a new Radeon Crimson Edition Beta, 16.7.1, which actually includes two fixes for the reported PCI-Express overcurrent issue that kept Internet people busy the last days.

The driver changelog mentions the following: "Radeon RX 480's power distribution has been improved for AMD reference boards, lowering the current drawn from the PCIe bus", and there's also a second item "A new "compatibility mode" UI toggle has been made available in the Global Settings menu of Radeon Settings. This option is designed to reduce total power with minimal performance impact if end users experience any further issues."

In order to adjust the power distribution between PCI-Express slot power and power drawn from the PCI-Express 6-pin power connector, AMD uses a feature of the IR3567 voltage controller that's used on all reference design cards.
This feature lets you adjust the power phase balance by changing the controller's configuration via I2C (a method to talk to the voltage controller directly, something that GPU-Z uses too, to monitor VRM temperature, for example). By default, power draw is split 50/50 between slot and 6-pin, this can be adjusted per-phase, by a value between 0 to 15. AMD has chosen a setting of 13 for phases 1, 2 and 3, which effectively shifts some power draw from the slot away onto the 6-pin connector, I'm unsure why they did not pick a setting of 15 (which I've tested to shift even more power).

The second adjustment is an option inside Radeon Settings, called "Compatibility Mode", kinda vague, and the tooltip doesn't reveal anything else either. Out of the box, the setting defaults to off and should only be employed by people who still run into trouble, even with the adjusted power distribution from the first change, which is always on and has no performance impact. When Compatibility Mode is enabled, it will slightly limit the performance of the GPU, which results in reduced overall power draw.

We tested these options, below you will find our results using Metro Last Light (with the card being warmed up before the test run). First we measured power consumption using the previous 16.6.2 driver, then we installed 16.7.1 (while keeping Compatibility Mode off), then we turned Compatibility Mode on.

As you can see, just the power-shift alone, while working, is not completely sufficient to reduce slot power below 75 W, we measured 76 W. As the name suggests, the changed power distribution results in increased power draw from 6-pin, which can easily handle slightly higher power draw though.
With the Compatibility Mode option enabled, power from the slot goes down to 71 W only, which is perfectly safe, but will cost performance.

AMD has also promised improved overall performance with 16.7.1, so we took a look at performance, using Metro again.
Here you can see that the new driver adds about 2.3% performance, which is a pretty decent improvement. Once you enable Compatibility Mode though, performance goes down slightly below the original result (0.8% lower), which means Compatibility Mode costs you around 3%, in case you really want to use it. I do not recommend using Compatibility Mode, personally I don't think anyone with a somewhat modern computer would have run into any issues due to the increased power draw in the first place, neither did AMD. It is good to see that AMD still chose to address the problem, and solved it fully, in a good way, and quick.
Add your own comment

147 Comments on AMD Releases PCI-Express Power Draw Fix, We Tested, Confirmed, Works

#1
Basard
So, if your mobo fries... simply reboot and turn compatibility mode on. Seems like it should be on by default.
Posted on Reply
#2
Chaitanya
nice of Amd to fix the power draw management.
Posted on Reply
#3
Countryside
Basard said:
So, if your mobo fries... simply reboot and turn compatibility mode on. Seems like it should be on by default.
It fries when you use the 480 with a ancient motherboard witch has an ancient power delivery system.
Posted on Reply
#4
cryohellinc
Countryside said:
It fries when you use the 480 with a ancient motherboard witch has an ancient power delivery system.
Better safe then sorry
Posted on Reply
#5
raptori
Basard said:
So, if your mobo fries... simply reboot and turn compatibility mode on. Seems like it should be on by default.
"the adjusted power distribution from the first change, which is always on" by default simply by installing the new driver .

The card will not draw full current while being idle on desktop , so obviously you can turn the second adjustment option (Compatibility Mode) on before running any heavy 3D application .
Posted on Reply
#6
Nobody99
Shouldn't motherboard deliver a max of 75W per lane, even if PCIe card requests more?
Posted on Reply
#7
ZoneDymo
well thats very neat.
So it actually works and the performance does not suffer :)

Nice job AMD.

@tPU and chance of running a few more games to see how they react?
Posted on Reply
#8
RejZoR
Told ya all that modern GPU's have clever power delivery systems and it was an easy fix by just redirecting power from PCIe to 6pin. This means that fix doesn't actually gimp the performance AT ALL because it redirects power to 6pin. Compatibility Fix switch however restricts it to initially advertised 150W.

So, AMD actually delivered best of both worlds. Not only they fixed power delivery of PCIe without performance penalty, they actually delivered factory over-boosted card at expense of slightly higher power consumption. Because when you flip Compatibility mode ON, you actually restrict it to official 150W.

So, for consumers, it's actually a double win. Good job AMD and good job reviewers spotting this anomaly. In the end, we, the consumers benefit the most from all this.

EDIT:
They should rename Compatibility Mode to Power Saving+, Enhanced Power Saving or something more descriptive by itself than Compatibility Mode which can mean anything from software features to hardware features...
Posted on Reply
#9
laszlo
Nobody99 said:
Shouldn't motherboard deliver a max of 75W per lane, even if PCIe card requests more?
mb don't have built in a current limiter and any consumer attached to it can draw more till something is kaboom
Posted on Reply
#10
Norton
Moderator & WCG-TPU Captain
My card is in a UPS truck atm so it will be waiting for me when I get home :)

Very nice of AMD to get the bugs worked out while I was waiting for the card to arrive :D
Posted on Reply
#11
mcraygsx
Basard said:
So, if your mobo fries... simply reboot and turn compatibility mode on. Seems like it should be on by default.
You simply cannot fry a mainboard when a minor spikes like that. Each PCI-E is able to handle that kind of load. If you fry your mainboard most likely it is not your 480 that caused it.

Original Post: Glad they were able to tweak and address the problem with a software update.
Posted on Reply
#12
RejZoR
It just hit me. If they could shift the power delivery, so can users. And expand it beyond total 150W/166W. The card has a very beefy VRM segment. And a 6pin which can deliver up to 150W on its own. Meaning this card has an overclocking potential if users can shift the power delivery and stick better cooling on them. Hidden massive potential!
Posted on Reply
#13
theoneandonlymrk
Sweet now if next payday would hurry up and get here I can get me some waterblocks and crank these to 11.
Posted on Reply
#14
3rold
Really nice to see good news from AMD. We need good participants in a "GPU Competition" and AMD is doing fine rn. I am waiting for custom RX 480 for my PC. I wanted to see what the 1060 fuss was all about but it seems to limited for future upgrades. Maybe the DirectX 12 Mixed Multi-GPU thing will be glorious, but it is still far from ready and I'm sure even if it becomes a thing, NVIDIA will just disable it just like it has done in the past.
Posted on Reply
#15
TRWOV
Basard said:
So, if your mobo fries... simply reboot and turn compatibility mode on. Seems like it should be on by default.
A single 480 won't fry your motherboard, even with the old driver. The only documented case of a damaged motherboard had multiple cards installed. If, for example, a board has 3 PCIe slots it has to have provisions for delivering 75w to each one, so 225w in total. A single 480 won't draw that much current from the motherboard.

The new driver sets the 12V bias to the 6pin connector, the compatibility mode just puts the power limit to 150w. It's a different fix for folks that still have problems with the 6pin bias fix; likely those with very old motherboards with power delivery systems that can't deal with the load spikes.


RejZoR said:
It just hit me. If they could shift the power delivery, so can users. And expand it beyond total 150W/166W. The card has a very beefy VRM segment. And a 6pin which can deliver up to 150W on its own. Meaning this card has an overclocking potential if users can shift the power delivery and stick better cooling on them. Hidden massive potential!
The bias has a maximum of .15 according to W1zard. You can't set all the power to come from the 6pin connector and even if you could it's not advisable anyway since each phase can deal 40w. If you switched everything to the 6pin it could only draw 120w at most.
Posted on Reply
#16
RejZoR
So, all phases can churn out 240W theoretical load and realistic max being 120W on 6pin phases and 75W on PCIe phases. That's still 195W which ain't that bad, giving some headroom. If you can cool it well enough.
Posted on Reply
#17
buildzoid
TRWOV said:
A single 480 won't fry your motherboard, even with the old driver.

The new driver sets the 12V bias to the 6pin connector, the compatibility mode just puts the power limit to 150w. It's a different fix for folks that still have problems (likely those with very old motherboards... we're talking about LGA775, AM2 boards).




The bias has a maximum of .15 according to W1zard. You can't set all the power to come from the 6pin connector and even if you could it's not advisable anyway since each phase can deal 40w. If you switched everything to the 6pin it could only draw 120w at most.
Each phase can do 40A not W and that's with the VRM at 125C. At normal operating temps (80C) you can shove much more current through them.
Posted on Reply
#18
zAAm
Nice, exactly the response we needed. Glad to see the VRM could handle per-phase load configuration.
Posted on Reply
#19
TRWOV
buildzoid said:
Each phase can do 40A not W and that's with the VRM at 125C. At normal operating temps (80C) you can shove much more current through them.
Ow, sorry, you're right. I recalled reading about them supplying 40 something... so watts seemed like a safe bet :)
Posted on Reply
#20
RejZoR
buildzoid said:
Each phase can do 40A not W and that's with the VRM at 125C. At normal operating temps (80C) you can shove much more current through them.
40A per phase, that's a... a lot more than 40W. At 12V, that's 480W per phase at 125°C. That can't be right. Tell me I've cocked up my math somewhere...
Posted on Reply
#21
zAAm
RejZoR said:
40A per phase, that's a... a lot more than 40W. At 12V, that's 480W per phase at 125°C. That can't be right. Tell me I've cocked up my math somewhere...
I think that would be 40A per phase on the secondary side, meaning 40A at the core/memory voltage, i.e. <1.5V.
Posted on Reply
#22
Shamalamadingdong
RejZoR said:
It just hit me. If they could shift the power delivery, so can users. And expand it beyond total 150W/166W. The card has a very beefy VRM segment. And a 6pin which can deliver up to 150W on its own. Meaning this card has an overclocking potential if users can shift the power delivery and stick better cooling on them. Hidden massive potential!
Here you go:
http://www.overclock.net/t/1604979/a-temporary-fix-for-the-excess-pci-e-slot-power-draw-for-the-reference-rx-480-cards/110#post_25324219

It increases the TDP to 250W~ by shifting the power targets, so -50% actually becomes the current/stock (0%) default of 110W and the +50% is now 250W. I believe these values are for the GPU/vCore itself and not the entire card. So you can imagine the added headroom.

An added bonus is that it also increases the limit on the memory so the maximum is 2500 MHz instead of 2250.

As always with mods like these: use at your own risk. Things can go kaboom if you don't know what you're doing.
Posted on Reply
#23
RejZoR
zAAm said:
I think that would be 40A per phase on the secondary side, meaning 40A at the core/memory voltage, i.e. <1.5V.
That would make more sense, 40A*1.5V is 60W. If that's a 125°C rating, at lower temperatures we can still be talking around 80-100W. That's still huge.

@Shamalamadingdong
Yeah, I'm aware of the risks, but that applies to every overclock. Still, the headroom can be a lot higher than around 150W as initially thought. If user has PSU powerful enough to feed 6pin, VRM is there to deliver it if needed. Far beyond the original specs. Meaning AIB's only have to slam better cooling on reference design. They don't even have to redesign PCB and components on it to gain headroom.
Posted on Reply
#24
GhostRyder
So its resolved then? Ok, guess we can knock that problem off the list...

Now then, we just need the aftermarket variants to allow better overclocking or better voltage control on the reference (I assume we will still get that at some point).
Posted on Reply
#25
xkm1948
I am having stability issue with this new driver on FuryX. 1100 OC is no longer stable no matter what amount of vcore offset is applied now. Played CSGO for 30mins and it crashed on me which have not happened since launch driver. Lower down to 1075OC and it still would crash. It can't even pass the 3DMark FireStrike Stress test now. I use DDU between all of my driver installations so it is definitely not the problem of driver files conflict. Reverted back 16.6.1 and everything is back to normal. Passed 100 loops of FSE Stress test at 99.5% for 1100 OC.

Very likely RTG is focusing on 480 driver now and somehow messed up driver support for previous gen cards. Guess I will just wait until a newer driver that does not solely focus on RX480 then.

And yes I have already submitted the problem to RTG.
Posted on Reply
Add your own comment