Saturday, July 2nd 2016

Official Statement from AMD on the PCI-Express Overcurrent Issue

AMD sent us this statement in response to growing concern among our readers that the Radeon RX 480 graphics card violates PCI-Express power specification, by overdrawing power from its single 6-pin PCIe power connector and the PCI-Express slot. Combined, the total power budged of the card should be 150W, however, it was found to draw well over that power limit.

AMD has had out-of-spec power designs in the past with the Radeon R9 295X2, for example, but that card is targeted at buyers with reasonably good PSUs. The RX 480's target audience could face troubles powering the card. Below is AMD's statement on the matter. The company stated that it's working on a driver update that could cap the power at 150W. It will be interesting to see how that power-limit affects performance.
"As you know, we continuously tune our GPUs in order to maximize their performance within their given power envelopes and the speed of the memory interface, which in this case is an unprecedented 8 Gbps for GDDR5. Recently, we identified select scenarios where the tuning of some RX 480 boards was not optimal. Fortunately, we can adjust the GPU's tuning via software in order to resolve this issue. We are already testing a driver that implements a fix, and we will provide an update to the community on our progress on Tuesday (July 5, 2016)."
Add your own comment

358 Comments on Official Statement from AMD on the PCI-Express Overcurrent Issue

#126
Secoya
$ReaPeR$
yeah mate.. whatever you say..

on topic: this doesnt seem to be such a major problem, and i love how people have blown it way out of proportion, what do you think happens when you oc a card geniuses?
Well, what normally happens is that if you have plenty of power source overhead to draw from, you never exceed the safe limit of it.

A 180W TDP GTX 1080 has a 225W power supply.

A 150W TDP GTX 1070 has a 225W power supply.

A 170W TDP RX 480 has a 150W power supply....

See the difference? The RX 480 was exceeding the limit without OCing too, which makes this "oversight" epic in terms of being bad for the consumer.

I'm certain those people who have smoke their motherboards really think this whole tihng has been blown out of proportion.

AMD RX 480 NERF incoming btw.
Posted on Reply
#127
cadaveca
My name is Dave
john_
AMD should know by now that even if one site finds something that it looks like a problem, all hell will get lose. If they where thinking like that, then maybe AMD's engineers live under a rock and have no contact with internet and the real world. Like Nvidia's engineers who where also living under a rock and never noticed that one of their company's products was selling for over 6 months with wrong specs. Or someone in AMD is a moron. Probably he also insists in giving exclusive interviews to unfriendly sites.
It's just simply possible the limit was "unlocked" to be controlled by driver, and the driver did not work right. Like, why offer power limit controls in driver in the first place? It only makes sense to me that the driver would control the power limit and prevent PCIe spec from being overrun without manual changes and some sort of pop-up disclaimer. I can't fathom any way that a card would exceed PCIe specs under any conditions, so I make up reasons why it might be acceptable.
Posted on Reply
#128
TheinsanegamerN
GC_PaNzerFIN
Reducing operating voltage MAY increase performance in thermally limited situations. Decreasing total power limit without changing anything is going to have opposite effect.
Reducing operating voltage MAY work on some GPUs, but due to obvious negative effect on stability (there is a reason why they put the VID it has now) it is too risky to do on all cards.
AMD has also been known in the past to put way to much voltage through their parts to increase yields in the past, especially for mid range high demand products.

Take the llano APUs. The mobile variants pulled 1.3V at their non turbo clock, and 1.415 volt for full boost clock. Seeing those things hit above the base clock was like seeing a unicorn.

Lo and behold, AMD used super relaxed settings to try and boost yields. Most, although not all, APUs could be undervolted. And I dont mean the -50mv that you get out of a mobile i7. You could typically get -350mv off of the core, while running at a much higher speed. For instance, my A6-3400m could do 2.1 GHz at 1.0375 volt, compared to 1.4GHz at 1.3 volt at stock. I could hit the 2.3 GHZ boost clock with 1.1 volt, compared with the 1.415 volt that the stock boost needed. And this was common, a huge number of llano chips acted this way, with only the rare model actually needing that much voltage to stay stable.

So it wouldnt surprise me if AMD could undervolt most 480s without difficulty. How they would do that in a driver is beyond me, but the headroom may be there.
Posted on Reply
#129
Dippyskoodlez
If this is a driver fixable issue, what about non official driver users?

That is NOT a good solution.
Posted on Reply
#130
RejZoR
What "non official" driver users? There is just one driver. The official one.

Like Cadaveca said, I think it was a driver cockup as well from the beginning and that's also one of reasons why cards actually perform kinda bad, because they go over the limit while they shouldn't be doing that. They were tested for stock operation and tuned for that. Including the fan profile. Card heats up more than it should for the factory fan profile, making it thermal throttle as well as hitting power limit.

Lets just wait for the damn promised fix and then evaluate it. Damn, AMD makes a statement and instead of people acknowledging it and waiting for the fix, they keep on dramatizing about it. Why we never see that for NVIDIA? Everyone bunch of fanboys and NVIDIA stock owners? Apparently...
Posted on Reply
#131
McSteel
HTC
I found this video to be quite informative:


As I have been saying... If AMD had been smart and made the card overdraw from the 6-pin PCI-E and NOT the MB slot, no problem at all, let it sip current.

The 5 +12V pins on the card and in the slot are simply not meant to carry that much.

This can only be fixed by having the VRM-in completely split, so that the PCI-E Slot and PCI-E External power are separated. Meaning, if there are 6 power phases, feed one, maybe two of them from the slot, and the rest from the external connector. This the only way to limit slot consumption.
Posted on Reply
#132
RejZoR
Well, or limiting it to actual 150W as they specified which is what they'll most likely do.
Posted on Reply
#133
BiggieShady
RejZoR
Everyone bunch of fanboys and NVIDIA stock owners?
By my calculation you'd have to invest 6.11 million dollars into nvidia to have yearly dividends at 70k usd so you could quit your job and spend all your time on forums deviously trying to raise nvdia stock price.
Posted on Reply
#134
RejZoR
Trust me, people do this sort of stuff for a lot less...
Posted on Reply
#135
john_
newtekie1
Except that isn't how it works on modern cards anymore. They have power limits in place to make sure they don't go over their power target. The power limit on the GTX950 with no power connector was 75w. You can up the clocks all you want, but GPU Boost will make them drop back down to keep within the 75w power limit. That is why tests like furmark don't give stupid high power numbers anymore. So overclocking without raising the power limit would still give ~75w.
In that review average power was 74W. After overclocking the GPU at over 1400Mhz and memory at over 2000MHz, the results where close to 20% extra performance. I think I am going to doubt that someone gets a 20% extra performance and stays under the 75W limit. Probably the card goes to 90W(20% extra power for 20% extra performance), if not more considering that usually power consumption goes faster up, compared to performance. If they where getting 1-3% extra performance, I would have agreed with you.
Interesting analogy, but somewhat fitting. In fact, there are MLC SSDs that perform crazy good, almost at SLC levels until you have them filled up a certain amount, then the performance starts to drop off. The reason being that they run all the MLC flash in SLC mode until the extra space is needed, then it switches to MLC mode. But the benchmarks in the review sure look good.
Are those SSDs advertised as SLC SSDs? I believe not. Well, if they where Nvidia products, probably they would. And people would be happy to convince themselves that while being MLC, performing as SLC would made them equal to SLCs. And anyone saying the opposite, would have been a stupid fanboy that hates Nvidia and doesn't acknowledge Nvidia's superior engineering. "It is a good design".

Also SLC vs MLC is not just performance difference. If I am not mistaken SLCs are considered as having better longevity. The same applies to the 970. It is not just those slow 500 MBs. Also less cache, less ROPs, less memory bandwidth. Specs where completely wrong and we shouldn't be giving any excuses to companies.

qubit
Just seen the whole video and now I'm even happier that I haven't bought an AMD card since 2008. Signifcant driver and performance glitches are one thing (and bad enough) but potentially killing the mobo with excess current is a new low. There's no way they couldn't have known about this at the design and testing phase. No, they tried to palm off a substandard product and hoped they wouldn't get cought out. It amounts to a kind of fraud FFS. :rolleyes:

IMO these cards should be pulled from the market until the fix has been applied and tested to be effective.

The way this company is going I'm unlikely to ever buy one of their graphics cards again. No wonder NVIDIA can charge what they like for their cards. At least they work beautifully most of the time.
Google Bumpgate + Nvidia. That's a low that AMD probably will never reach. Also I bet you haven't downloaded a single Nvidia driver the last 12 months, considering that you still talk about drivers. Well I think McCoy's words about this argument would have been "It's dead Jim".

McSteel
As I have been saying... If AMD had been smart and made the card overdraw from the 6-pin PCI-E and NOT the MB slot, no problem at all, let it sip current.
I guess they thought that there are just too many "600W PSUs" costing $20 out there.

What AMD should have done was to lock the GPU at a specific frequency and say "Sorry guys, you will have to buy a custom if you want overclocking". Or they could just offer only a 4GB reference version at $199 and let AIBs made the 8GB cards. 4GB less GDDR5 on board could also help in lowering power consumption.
Posted on Reply
#136
BiggieShady
So why does this happen ... stumbled upon the twitch video from @buildzoid (kudos dude) at 53:40 minutes starts the "gotcha" moment where insane board power delivery design is explained ... out of spec 6 pin connector used as an 8 pin and yet half of vcore vrms connected to pcie power pins.

The video: https://www.twitch.tv/buildzoid/v/75850933?t=53m40s
Posted on Reply
#137
$ReaPeR$
Secoya
Well, what normally happens is that if you have plenty of power source overhead to draw from, you never exceed the safe limit of it.

A 180W TDP GTX 1080 has a 225W power supply.

A 150W TDP GTX 1070 has a 225W power supply.

A 170W TDP RX 480 has a 150W power supply....

See the difference? The RX 480 was exceeding the limit without OCing too, which makes this "oversight" epic in terms of being bad for the consumer.

I'm certain those people who have smoke their motherboards really think this whole tihng has been blown out of proportion.

AMD RX 480 NERF incoming btw.
how many are affected? find me the number. we are talking about 16watt over the specs, 16 ffs.

[IMG]http://www.guru3d.com/index.php?ct=articles&action=file&id=23242[/IMG]
Posted on Reply
#138
RejZoR
Sure. Lets go by the book about the specs. So, where do Titan-Z and R9 295X2 get power from then if 8pin is specified up to 150W and it has 2 of them? That's 300W + 75W on PCIe slot. Where's the rest of 110W+ coming from then? From thin air?

Apparently if RX480 would draw those extra 16W from 6pin all would be pink and fluffy. But PCIe oh noes everyone running around losing their shit. Ever thought the fix might involve just that? Limiting power to actual 150W or drawing more power from 6pin? But no, lets generate even more unnecessary drama. When NVIDIA fucks up, everyone gets defensive to stupendous levels. AMD fucks something up, everyone loses their shit and creates so much drama around it even Venezuelan soap operas look shy in comparison... Por favor!
Posted on Reply
#139
zAAm
RejZoR
Sure. Lets go by the book about the specs. So, where do Titan-Z and R9 295X2 get power from then if 8pin is specified up to 150W and it has 2 of them? That's 300W + 75W on PCIe slot. Where's the rest of 110W+ coming from then? From thin air?

Apparently if RX480 would draw those extra 16W from 6pin all would be pink and fluffy. But PCIe oh noes everyone running around losing their shit. Ever thought the fix might involve just that? Limiting power to actual 150W or drawing more power from 6pin? But no, lets generate even more unnecessary drama. When NVIDIA fucks up, everyone gets defensive to stupendous levels. AMD fucks something up, everyone loses their shit and creates so much drama around it even Venezuelan soap operas look shy in comparison... Por favor!
Actually, drawing extra power from the 6-pin or 8-pin PCI-E connector should be relatively safe. Drawing extra power from the motherboard however is not. If you limit your power to 75W for the motherboard you can draw a lot of "out of spec" power from the PCI-E power connector if your PSU is beefy enough. The connectors and accompanying 18AWG wires should be fine for more than the PCI-E rated power (probably close to double if you look at the actual connector datasheets). However, motherboard power tracks and the small contacts of the PCI-E connector aren't designed with that sort of overhead in mind. So AMD could get away with cheating the spec as long as they don't exceed it on the motherboard side. If they can put out a software fix that only limits the motherboard connected phases to 75W then all should be well without actually reducing the TDP.
Posted on Reply
#140
BiggieShady
zAAm
If they can put out a software fix that only limits the motherboard connected phases to 75W then all should be well without actually reducing the TDP.
Well, it seems some of the VRM phases are connected directly to the pcie power pins ... if that's true software fix is unlikely to happen short of undervolt and underclock ... See my post above
Posted on Reply
#141
john_
zAAm
Actually, drawing extra power from the 6-pin or 8-pin PCI-E connector should be relatively safe.
ACE 600W Black ATX Gaming PC PSU Power Supply 120mm Red | eBay

600W, £17.24 inc. VAT

I think there are more (quantity) dangerous PSUs out there, than motherboards. And those PSUs will not go alone to the afterlife. They will take other parts of the hardware with them. Probably the motherboard too. This is the only excuse for AMD choosing the PCIe bus over the 6pin power connector I can think of. Of course I still believe that they have no excuse for the whole mess.
Posted on Reply
#142
RejZoR
But NVIDIA had one (excuse) when they fucked it up on several occasions in the past? Oh boy the double standards...
Posted on Reply
#143
buggalugs
This is a beatup., Lots of hardware runs out of spec. Lots of graphics cards run out of spec. There is enough headroom built into the spec to handle this.

Its kind of strange how people are going crazy about this, but those same people overclock the crap out of their parts and run out of spec. How do you think we can overclock our shit without our computers blowing up?? Because the hardware is designed to handle more than the spec allows.

This is a non-issue for 99.9% of people unless you have a crappy cheap motherboard from 2005. Any graphics card update can stress an old motherboard, same with other parts like a PSU. A motherboard is most at risk of dying after a major hardware upgrade when it is old.

AMD are releasing a driver update anyway just to shut people up.. If I had a 480 I wouldnt want the driver update, I'd prefer they leave it alone.
Posted on Reply
#144
HTC
jabbadap
Sad part is, they should not have to even make it worse. They could have restrict pcie slot power below the spec and take over power from 6-pin connector(best practice, this way you don't over power pcie slot even while overclocking) or slap that damn 8-pin connector to it.
$ReaPeR$
how many are affected? find me the number. we are talking about 16watt over the specs, 16 ffs.

[IMG]http://www.guru3d.com/index.php?ct=articles&action=file&id=23242[/IMG]
How many of these exceed the PCI-e portion of the TDP? I'm not talking about the whole wattage: only the PCI-e part of it.

Ordinarily, you'd look @ the PSU for higher wattage needs but, with this card, you can have a ... say ... 1000W PSU and STILL have problems simply because the PCI-e is using more power then it should. This could end up with the motherboard's PCI-e slot's contacts burned because, unlike the 6/8 pin connectors, they are NOT made to overclock.
Posted on Reply
#145
RejZoR
I don't remember, but is the actual PCIe drawing this much? Sure, it's 166W, but is it actually from PCIe or is it over 6pin? Everyone seems to just assume 6pin is absolutely strict 75W so it ha to be PCIe then. But is it? Who has actually measured it at PCIe? Can't remember the testers who would do this at the moment...
Posted on Reply
#146
laszlo
the 6 pin connector has pin #2 also delivering 12 V so this 6 pin connector can deliver more than 75W from PSU up to 150W as i see ; i don't know if rx480 PCB has pin 2 connected physically to use current from it....

i'm expecting the release of aib with 8 pin to be sure that i'll buy a correct card.
Posted on Reply
#147
HTC
RejZoR
I don't remember, but is the actual PCIe drawing this much? Sure, it's 166W, but is it actually from PCIe or is it over 6pin? Everyone seems to just assume 6pin is absolutely strict 75W so it ha to be PCIe then. But is it? Who has actually measured it at PCIe? Can't remember the testers who would do this at the moment...
https://www.techpowerup.com/forums/threads/official-statement-from-amd-on-the-pci-express-overcurrent-issue.223833/page-5#post-3482710

It's BECAUSE it's drawing more then it should that this is a problem to begin with.

Let me give an example. Look @ 2 cards that draw 180W:

- card A: 70W from PCI-e + 110W from 6 pin connector
- card B: 85W from PCI-e + 95W from 6 pin connector

Both cards are out of spec BUT one can end up burning your motherboard (worst case scenario) while the other should not: can you tell which one?
Posted on Reply
#148
TRWOV
qubit
Just seen the whole video and now I'm even happier that I haven't bought an AMD card since 2008. Signifcant driver and performance glitches are one thing (and bad enough) but potentially killing the mobo with excess current is a new low. There's no way they couldn't have known about this at the design and testing phase. No, they tried to palm off a substandard product and hoped they wouldn't get cought out. It amounts to a kind of fraud FFS. :rolleyes:

IMO these cards should be pulled from the market until the fix has been applied and tested to be effective.

The way this company is going I'm unlikely to ever buy one of their graphics cards again. No wonder NVIDIA can charge what they like for their cards. At least they work beautifully most of the time.
nVidia cards also do that but somehow AMD always gets the drama for some reason... well, to be fair the reason in this case is that AMD knew the card would pull >150w at times and they should have shipped with an 8pin connector but choose to go with a 6pin instead for marketing reasons (the worst kind of reasons).


Thankfully the fix is somehow easy but it's going to be a two part fix as far as I can tell seeing the quoted videos.

1) The fix AMD is going to push via drivers. This will likely set the power limit to 150w and call it a day. Not different from setting it yourself on Wattman.

2) A bios update to curb the PCIe power delivery to 75w and let the rest of the power come from the PCIe slot. The VRM phases are capable of outputting 40w each so there should be no issue for doing this.

The second one is trickier but AMD could potentially release and automatic bios flash tool and label it as a beta driver or hotfix or something.


Why this happened in the first place? I would like to know if all the cards have the same bios version on them. AMD says that the card shouldn't be behaving that way so I have the feeling that there are some 480s out there with a 50/50 power split bios.
Posted on Reply
#149
RejZoR
Because NVIDIA fanboys are more fanboyish than the rest...
Posted on Reply
#150
okidna
RejZoR
I don't remember, but is the actual PCIe drawing this much? Sure, it's 166W, but is it actually from PCIe or is it over 6pin? Everyone seems to just assume 6pin is absolutely strict 75W so it ha to be PCIe then. But is it? Who has actually measured it at PCIe? Can't remember the testers who would do this at the moment...
PCPer, they did a stock analysis, increasing the power limit analysis, and even debunking an analysis about GTX 960 STRIX : http://www.pcper.com/reviews/Graphics-Cards/Power-Consumption-Concerns-Radeon-RX-480
Posted on Reply
Add your own comment