AMD once again violating power specifications? (AMD RX-480)

[laminutederire]

3 hours ago, Megazero said:

acknowledge this problem and said that a simple driver issue wont fix it, so this is fact.

When have they talked about the driver fix not working?

[LAwLz]

42 minutes ago, Ryan_Vickers said:

 

So the conclusion I was looking for earlier is something like this:

• Yes, the 480 breaks specs in a pretty big way.  Its constant loads are above spec, not just the spikes
• Yes, this is a problem
• No, it's not just a problem because AMD does it
• Other cards (yes Nvidia cards) have had issues too but not like this.  The ASUS 960 was the worst example we can come up with and that only spiked above for short durations.  Yes, this is a difference that matters.

Yep, exactly.

I would like to add two more points as well.

 

• We don't know what the effects will be like. It might be that it damages your motherboard or other components, it might be that it causes stuttering in games (since the GPU might not get enough power), it might be that it causes compatibility issues with other add-in cards, it might cause audio issues... The list goes on. PCPer is going to test it on a low-end but modern motherboard to see if something breaks. The problem with that test is that it will not show any long term effects it might have. Like for example it might kill components slowly over a year or several years.
• Custom PCB versions of the RX 480 will be able to fix this issue by using 8-pin power connectors and/or better load balancing.
• We do not know if this can be fixed via a software update or not. If it can then we don't know what other effects it might have (worse performance because it will power throttle?).

 

19 minutes ago, mariushm said:

Motherboards have to provide those 5.5A at 12v to EACH pci-e x16 slot on them, so they are designed with large copper fills and traces going to the pci-e slots area, they are capable of providing those watts to the slots. The motherboards won't really care that the RX 480 pulls an average of 80 watts instead of the maximum recommended in the standard of 12v x 5.5a = 66 watts.

I see a lot of people use the 80 watt number from PCPer and I think it is important to note that the 80 watts in Ryan's test was for the +12V rail alone, which is only rated for 66 watts (with 8% safe-guard). (I know you mentions it in your post but I want everyone to fully understand what is going on).

So when you say 80 watts it might sound like it's only 5 watts over the spec, but it's actually 14 (21% above the spec).

On top of that, their test was a general gaming session. So during your average gaming scenario it is 21% above the spec. If you do anything that stresses the GPU even more (basically any GPGPU) then it will be even more.

That's why Ryan described their numbers as "bad case scenario by by no means a worse case scenario".

 

19 minutes ago, mariushm said:

The extra constant power consumption also shouldn't be too much of an issue because it's not that much extra, any good motherboard will handle it just fine. It would have been a big deal if the constant power consumption would have been let's say 100 watts or more. 

Well according to the PCPer article where they talked to someone in the motherboard industry, their "bad case" scenario will likely cause damage over time.

 

 

3 minutes ago, laminutederire said:

When have they talked about the driver fix not working?

JayzTwoCents said it couldn't be done and he has been talking with AMD apparently. I wouldn't take that as a fact though.

I mean, in a worst case scenario they could limit the frequency. That can be done though a driver and that will definitely solve the issue.

The RX 480 scales horribly with frequency (at least over the stock frequency) so even a fairly small decrease in frequency would drive the power consumption down quite a lot.

[mariushm]

2 minutes ago, Megazero said:

Isnt the test from Tom say the average draw from PCIE of the 480 is 100W?

Also really? the fix is to underclock the card? Why should anyone buy a card and HAVE to underclock it to make it work? 

Another problem with your statement is "good motherboard", how do you define "good"? Remember who this card is targeting at: budget gamer. Those that spend only $200-$240 on a GPU is likely to have a cheap Mobo too. They ofc wont pair this with say... at $150+ mobo. So this make this problem even worse.

Because this card apparently has half it's VRM connected to the motherboard power and half is connected to the pci-e 6 pin connector, it's only natural that it will draw relatively equal amounts of power from both power sources. Therefore, when they overclocked the card making it use up to 200 watts in total, it's kind of obvious the card pulled an average of 100 watts from each power source.

Yes, this is bad, but the card does not do this by default, in the default configuration the card is sold to you. You only come up to this by overclocking the card way above the default 1266 Mhz they sell the card with. You assume all the risks when you're doing overclocking.

 

Perhaps the issue is that most overclockers assume the video cards must have electronic circuitry to limit themselves from pulling more than the recommended value from the slot and this card apparently doesn't have such circuitry.

 

I consider a "good" motherboard one with the latest chipsets from AMD, not motherboards that use obsolete chipsets to save pennies, for example motherboards with the SB710 chipset and with 4 phase VRMs that can only handle up to 95w TDP processors and so on. Basically, I'd say any $70+ motherboard should handle a card like RX 480 just fine.

 

No, I don't think it's right for AMD to neuter the card by reducing the 1266 Mhz boost frequency. But I think that's what the software fix will do. Another simple fix would be I guess to add in the firmware some code that measures how much time the card stays at 1266 Mhz and if it stays for more than x ms , it will then artificially reduce it by a few tens of Mhz for a few ms and then bring it back, so that over time, the average consumption drops.

You'd get upset if AMD reduces the boost frequency. OK, so don't reduce it, leave the value there but make the card achieve that number less often.

 

But on the other hand, let's look at cards like GTX1080 founder's edition. Do you prefer that behavior, where the card has a boost frequency of 1866 Mhz but which practically is only achieved for a few minutes and then the card overheats and throttles itself down to 17xx frequencies? Isn't this kind of the same thing except instead of power we have temperature?  nVidia is also "lying" with that card, selling you something that says 1866 Mhz but which in practice is only achieved for small periods of time.

 

Why would reducing this maximum boost frequency or reducing the percentage of time the card can reach this 1266 Mhz be bad then, wouldn't it be the same thing?

 

[Valentyn]

5 hours ago, Ryan_Vickers said:

 

So in one game the 480 falls short of the 970 and in another it's basically at 980 ti performance?  Hm...

It's performance is all over the place, even when only looking at DX11 or just DX12.

In GTA V the 480 is around 290 performance, yet in Fallout 4 with full Gameworks settings it's beating the Fury.

Seems like the drivers are still wonky, and that we need AIB cards with better designed cooling and power delivery.

[laminutederire]

2 minutes ago, LAwLz said:

 

The RX 480 scales horribly with frequency (at least over the stock frequency) so even a fairly small decrease in frequency would drive the power consumption down quite a lot.

I don't think that it scales badly, it's just that it power throttles at that frequency, therefore there is no significant performance change until it doesn't power throttle anymore.

Okay I ask since I just woke up after 8h ours of sleep and I could have missed AMD announcement. 

[zMeul]

1 hour ago, Ryan_Vickers said:

Exactly.  So, not only should we go beyond checking the total power consumption of the card against the sum of all suppliers by breaking it down by PCIe cable power and PCIe slot power, but we should break down the slot power into the different rails.  Has anyone tested it that closely?

Tom's already does this since they started measuring PEG power draw, years ago

PCPer started doing PEG power measuring recently

[zMeul]

53 minutes ago, Ryan_Vickers said:

Other cards (yes Nvidia cards) have had issues too but not like this.  The ASUS 960 was the worst example we can come up with and that only spiked above for short durations.  Yes, this is a difference that matters.

that's not an nVidia card - ASUS has splaining to do in that case

 

to my knowledge, the worst nVidia offender is a GTX750Ti that has no additional PSU connector, draws power only from the PEG

[vanished]

5 minutes ago, Valentyn said:

It's performance is all over the place, even when only looking at DX11 or just DX12.

In GTA V the 480 is around 290 performance, yet in Fallout 4 with full Gameworks settings it's beating the Fury.

Seems like the drivers are still wonky, and that we need AIB cards with better designed cooling and power delivery.

I wonder if it could be related to what was talked about in that video where he detailed how this card mainly power throttles over anything else.  Just as there are "power viruses" like furmark and prime95, perhaps some games draw a lot more power for the computations they do than others.  If that's the case, I'd expect these wild variations in performance to disappear as soon as we can get a card that's properly fed with power.

[Kobz360]

6 minutes ago, LAwLz said:

We don't know what the effects will be like. It might be that it damages your motherboard or other components, it might be that it causes stuttering in games (since the GPU might not get enough power), it might be that it causes compatibility issues with other add-in cards, it might cause audio issues... The list goes on. PCPer is going to test it on a low-end but modern motherboard to see if something breaks. The problem with that test is that it will not show any long term effects it might have. Like for example it might kill components slowly over a year or several years.

 

When this guy ran the 480 with his (very) old AM2 motherboard, the entire PC shut down at specific points in benchmarks (presumably the most graphically intensive scenes). He put a 980 Ti in after and the entire PC was fine. 

 

That's a pretty extreme scenario though - need someone to test it with like a cheap Asrock board or something, though I honestly doubt it'll be an issue. I'd be expecting to hear multiple complaints by now of people's PCs bricking and boards crashing but time will tell 

 

 

[vanished]

2 minutes ago, zMeul said:

that's not an nVidia card - ASUS has splaining to do in that case

 

to my knowledge, the worst nVidia offender is a GTX750Ti that has no additional PSU connector, draws power only from the PEG

The ASUS Strix GTX 960 is not an nvidia card?

[Megazero]

2 minutes ago, mariushm said:

Because this card apparently has half it's VRM connected to the motherboard power and half is connected to the pci-e 6 pin connector, it's only natural that it will draw relatively equal amounts of power from both power sources. Therefore, when they overclocked the card making it use up to 200 watts in total, it's kind of obvious the card pulled an average of 100 watts from each power source.

Yes, this is bad, but the card does not do this by default, in the default configuration the card is sold to you. You only come up to this by overclocking the card way above the default 1266 Mhz they sell the card with. You assume all the risks when you're doing overclocking.

 

Perhaps the issue is that most overclockers assume the video cards must have electronic circuitry to limit themselves from pulling more than the recommended value from the slot and this card apparently doesn't have such circuitry.

 

I consider a "good" motherboard one with the latest chipsets from AMD, not motherboards that use obsolete chipsets to save pennies, for example motherboards with the SB710 chipset and with 4 phase VRMs that can only handle up to 95w TDP processors and so on. Basically, I'd say any $70+ motherboard should handle a card like RX 480 just fine.

 

No, I don't think it's right for AMD to neuter the card by reducing the 1266 Mhz boost frequency. But I think that's what the software fix will do. Another simple fix would be I guess to add in the firmware some code that measures how much time the card stays at 1266 Mhz and if it stays for more than x ms , it will then artificially reduce it by a few tens of Mhz for a few ms and then bring it back, so that over time, the average consumption drops.

You'd get upset if AMD reduces the boost frequency. OK, so don't reduce it, leave the value there but make the card achieve that number less often.

 

But on the other hand, let's look at cards like GTX1080 founder's edition. Do you prefer that behavior, where the card has a boost frequency of 1866 Mhz but which practically is only achieved for a few minutes and then the card overheats and throttles itself down to 17xx frequencies? Isn't this kind of the same thing except instead of power we have temperature?  nVidia is also "lying" with that card, selling you something that says 1866 Mhz but which in practice is only achieved for small periods of time.

 

Why would reducing this maximum boost frequency or reducing the percentage of time the card can reach this 1266 Mhz be bad then, wouldn't it be the same thing?

 

You know the "latest chipset" from AMD is the age old AM3+ right? For Intel maybe, but again, lots of people have old mobo and would make up a large chunk of the people who only willing to spend $200-240 on upgrading to the 480.

AMD on the other hand that boosted clock advertised on the box is what causing this problem. And one thing to note is what Jay talk about: Due to this violate the PCIE standard and stuff, AMD have to fix it if they want to keep the "PCIE" logo on their box, either by recall or whatever, else they will face a lawsuit. 

As for comparing to Nvidia, yes I hate that it throttle too (not that I experiences it at all since... AIB card). But between the 2 I'd rather my card boosted and throttle by itself than having my mobo die or facing problem or auto shutdown upon reaching that wattage limit.

[Valentyn]

1 minute ago, Ryan_Vickers said:

I wonder if it could be related to what was talked about in that video where he detailed how this card mainly power throttles over anything else.  Just as there are "power viruses" like furmark and prime95, perhaps some games draw a lot more power for the computations they do than others.  If that's the case, I'd expect these wild variations in performance to disappear as soon as we can get a card that's properly fed with power.

I think that's quite possible, OC3D noted that if you increase the power limit by 30% the card will pull an extra 30% right away. Even if the clocks aren't increased either.

Although so far it seems AMD's new card does handle tessellation better than other cards, and is quite good in DX12, it just need to get a better revision out. That'll be the AIB versions. Will be interesting to see how they fare compared to the reference card.

[Prysin]

@LAwLz

actually, i must correct you a little bit

 

Stuttering will only occur if the GPU is dropping frequencies. But according to Tiny Tom Logan, who did a review on the power issues themselves, the GPUs (all of them) are POWER STARVED OUT OF THE BOX.

 

The XFX OC edition of the "stock" cards are clocked at 1.29GHz, whilst the rest is 1.15?. Anyway, his tests saw them score the same in Firestrike. But when he upped the power limit of the card, by 20-30%, all the GPUs shot up from 16.5k points, to 16800-17000. Just like that.

 

He talked with AMD about why the GPUs were power starved, and it had to do with heat.

They nerfed potential performance because heat escalates enourmously at a certain amount of wattage. More then the stock cooler can handle.

So in order to make the card more silent, and the GPU cooler, they kneecapped the RX 480 by power limiting it.

 

Actual performance of all the GPUs can be expected to be 10% higher or more with good aftermarket coolers. Which is a HUGE jump.

 

 

Also, AMD has officially stated that the RX 480 GPU DIE uses 110w. The board with memory, leakage, heat etc adds another 40w.

We know that GDDR5 is massively power hungry. I think that by reducing the memory clock from 8GB/s to 6.5-7GB/s they will be able to cut A LOT of power. As both the IMC but also the VRAM package itself will use MUCH less power.

 

it just may be enough to get it down to reasonable levels.

[zMeul]

18 minutes ago, Ryan_Vickers said:

The ASUS Strix GTX 960 is not an nvidia card?

the STRIX is an ASUS custom PCB with a GTX960 GPU on it - it's not nVidia's fault ASUS designed a out of spec PCB and power delivery circuitry

[vanished]

3 minutes ago, zMeul said:

the STRIX is an ASUS custom PCB with a GTX960 GPU on it - it's not nVidia's fault ASUS designed a out of spec PCB and power delivery circuitry

Of course  But in the "this card sux BURN NVIDIA" vs "this card sux BURN AMD" war, you can't expect people to see beyond that

[laminutederire]

3 minutes ago, zMeul said:

the STRIX is an ASUS custom PCB with a GTX960 GPU on it - it's not nVidia's fault ASUS designed a out of spec PCB and power delivery circuitry

And @Ryan_Vickers

Stop arguing on that, we already have pages on that.

If you don't like that example there us the 750Ti which was even worse than the RX 480, but anyway that doesn't matter we all know the bias against amd in some community etc, but that doesn't solve the problem.

[vanished]

3 minutes ago, laminutederire said:

And @Ryan_Vickers

Stop arguing on that, we already have pages on that.

If you don't like that example there us the 750Ti which was even worse than the RX 480, but anyway that doesn't matter we all know the bias against amd in some community etc, but that doesn't solve the problem.

Didn't mean to start that up again.  It just had to be a point in my conclusion for completeness (and fairness) that other cards, even on the "great and infallible green team" have had this happen, since, as you pointed out, there was a whole argument on that... actually this is quite ironic now that you mention it, I was trying to put everything to rest  

[zMeul]

5 minutes ago, laminutederire said:

And @Ryan_Vickers

Stop arguing on that, we already have pages on that.

If you don't like that example there us the 750Ti which was even worse than the RX 480, but anyway that doesn't matter we all know the bias against amd in some community etc, but that doesn't solve the problem.

if you actually bothered to read the god damn posts, and you didn't !!!, I said GTX750Ti is nVidia's worst offender

but nooo .. blame ne for fanboism, bias and general hate towards AMD - that will absolutely cure the RX480 

 

I remember when I made a joke about AMD ending up conducting their business from a trailer park, and mods + Linus removed my topic

the fucking irony!

[laminutederire]

2 minutes ago, zMeul said:

if you actually bothered to read the god damn posts, and you didn't !!!, I said GTX750Ti is nVidia's worst offender

I know you did, that's why I gave it back to you to say it's pointless to argue on the "nvidia vs amd" battle.

There was already a pretty lengthy one on that matter yesterday, and it doesn't need to continue.

 

4 minutes ago, Ryan_Vickers said:

Didn't mean to start that up again.  It just had to be a point in my conclusion for completeness (and fairness) that other cards, even on the "great and infallible green team" have had this happen

Yeah sure sorry, nvidia and intel fuck up as well as AMD, people prefer to forget it, but hey, some people know better

 

[LAwLz]

38 minutes ago, laminutederire said:

I don't think that it scales badly, it's just that it power throttles at that frequency, therefore there is no significant performance change until it doesn't power throttle anymore.

Okay I ask since I just woke up after 8h ours of sleep and I could have missed AMD announcement. 

It scales horrendously bad. I wasn't talking about performance. I was talking about frequency:watts.

PCPer applied a 3.5% overclock and the power consumption increased by 33% and then it throttles.

Tom's Hardware did not even report their overclocking numbers because the power consumption was already out of control.

 

That seems to indicate that lowering the frequency a bit would decrease power consumption quite a lot.

 

 

 

31 minutes ago, Kobz360 said:

That's a pretty extreme scenario though - need someone to test it with like a cheap Asrock board or something, though I honestly doubt it'll be an issue. I'd be expecting to hear multiple complaints by now of people's PCs bricking and boards crashing but time will tell 

PCPer will test it with some H170 based board. Didn't specify which one though.

 

 

 

24 minutes ago, Prysin said:

Stuttering will only occur if the GPU is dropping frequencies. But according to Tiny Tom Logan, who did a review on the power issues themselves, the GPUs (all of them) are POWER STARVED OUT OF THE BOX.

 

The XFX OC edition of the "stock" cards are clocked at 1.29GHz, whilst the rest is 1.15?. Anyway, his tests saw them score the same in Firestrike. But when he upped the power limit of the card, by 20-30%, all the GPUs shot up from 16.5k points, to 16800-17000. Just like that.

Stuttering can most definitely happen when the GPU does not get the correct power.

Try plugging your 295X2 into a ~30 dollar Fireball™ power supply and you will most likely see issues arise.

It might be that the GPU drops in frequency when it doesn't get the correct power (I haven't tested that) so you might be right, but our points can both be correct at the same time.

 

Tiny Tom Logan's test does not account for low-end motherboards and issues that might appear long term (such as the higher power draw wearing out components on the motherboard faster).

 

24 minutes ago, Prysin said:

He talked with AMD about why the GPUs were power starved, and it had to do with heat.

They nerfed potential performance because heat escalates enourmously at a certain amount of wattage. More then the stock cooler can handle.

So in order to make the card more silent, and the GPU cooler, they kneecapped the RX 480 by power limiting it.

That sounds very plausible since the overclocking numbers are so appalling. It makes me worried about how good cards with custom coolers will be though. It sounds like the GPU core is just not very good. So there is a risk that even a custom power delivery + better cooler still won't result in that much higher performance, if the threshold where heat and power spirals out of control is not that far above the stock frequency.

But that remains to be seen.

 

24 minutes ago, Prysin said:

Also, AMD has officially stated that the RX 480 GPU DIE uses 110w. The board with memory, leakage, heat etc adds another 40w.

No, they stated that the TDP is 110W for the GPU core, and the other components has a total TDP of 40W. TDP and power consumption are two separate things, and as we can clearly see the RX 480 uses more than it is rated for.

 

 

 

9 minutes ago, laminutederire said:

And @Ryan_Vickers

Stop arguing on that, we already have pages on that.

If you don't like that example there us the 750Ti which was even worse than the RX 480, but anyway that doesn't matter we all know the bias against amd in some community etc, but that doesn't solve the problem.

I don't get why people keep bringing up the 750 Ti. Here are the Tom's Hardware measurements for that card:

 

Do you understand the difference between peaks and average? If you do then you should understand why what the RX 480 does is bad, and what the 750 Ti does isn't.

 

 

 

10 minutes ago, Ryan_Vickers said:

Of course  But in the "this card sux BURN NVIDIA" vs "this card sux BURN AMD" war, you can't expect people to see beyond that

Please don't even joke about it. 99% of people talking about this right now has absolutely no idea what they are talking about, and will leech on and parrot whatever information they find that fits their predefined conclusion.

You have to be extremely careful to not say anything wrong, even as a joke, because idiots will not only believe it, but also spread it to all the other idiots.

[vanished]

3 minutes ago, laminutederire said:

I know you did, that's why I gave it back to you to say it's pointless to argue on the "nvidia vs amd" battle.

There was already a pretty lengthy one on that matter yesterday, and it doesn't need to continue.

 

Yeah sure sorry, nvidia and intel fuck up as well as AMD, people prefer to forget it, but hey, some people know better

 

My main point was that 

1. not every green card is perfect either
2. the 960 wasn't nvidia's fault though - it was ASUS, as mentioned
3. even the ASUS 960 wasn't as bad as this though, since although it did spike out of spec, at least the averages were OK

@LAwLz I see you're comment just came in, so this is sort of a reply to your quote of me too

Edited July 1, 2016 by Ryan_Vickers

[zMeul]

3 minutes ago, LAwLz said:

PCPer will test

PcPer conducted their tests on a ASUS Rmapage V if I recall right

[Guest Kloaked]

11 minutes ago, laminutederire said:

And @Ryan_Vickers

Stop arguing on that, we already have pages on that.

If you don't like that example there us the 750Ti which was even worse than the RX 480, but anyway that doesn't matter we all know the bias against amd in some community etc, but that doesn't solve the problem.

What about the 750 Ti? The reference designs were drawing within the amount they were supposed to from the PCIe slot.

And everyone has a bias. That's life.

[zMeul]

4 minutes ago, LAwLz said:

TDP and power consumption are two separate things

not when talking about ASIC, and the GPU is an ASIC - so yes, when AMD said the TDP for the ASIC(GPU) is 110W they were correct

almost ... becasue it was discovered the ASIC (GPU) draws ~130W avg - observed with GPU-Z and confirmed by TPU programmer who coded GPU-Z

[laminutederire]

4 minutes ago, LAwLz said:

 

Ah you're talking about that scaling! Then okay sure.

As for the 750Ti, I couldn't care less, it doesn't matter what it does, we're talking about what happens with the 480 and how it's going to be fixed, so I didn't informed me very well on that matter, since it's off topic anyway.