Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Don't buy the recent reference cards
There are really several intersecting thoughts here. I don't think I can structure this post without burying the lede somehow, so let's get the main thoughts up front.
1. Whatever happened to PowerTune?
2. Why does the reference RX 480 only have a single 6-pin PCI-E connector?
3. What if a driver update reduces clock speeds?
4. It's a good thing that third-party board partners are around.
Returning to the title, I'd include the GeForce GTX 1080, GeForce GTX 1070, and Radeon RX 480 in that. For the GeForce cards, it's really just a question of price. Do you really want to pay $700 (or $800 or $900, depending on how much you're gouged) for a product that you know will soon be $600?
But with the RX 480, it's something much worse. At least Nvidia was up front about pricing, if not timing. And the delays were incandescently obvious to those who understood the tech. But I've complained enough about Pascal, so I want to spend most of this post going after Polaris.
Some reviews noticed the Radeon RX 480 pulling 160 W. Now, there's nothing wrong with a desktop video card burning 160 W. But there's something very wrong with a PCI Express card with only a single 6-pin PCI-E power connector burning 160 W.
The PCI Express slot is rated as being able to deliver 75 W, a 6-pin connector also 75 W, and an 8-pin connector 150 W. If all you've got is the slot and a single 6-pin, that's 75+75 = 150 W. Pulling 160 W through that is running something out of spec.
Now, the Radeon RX 480 isn't the first card to do this. The GeForce GTX 470 had only two 6-pin PCI-E power connectors, and it routinely pulled more than 225 W.
No, I didn't just say that Polaris is as bad as Fermi. But I suspect that the reasons are the same: someone decided late in the game that the stock clock speed needed to increase. Rather than going back to the drawing board to beef up power delivery and make a card that could handle it, they just took the cards they had and clocked them higher.
And the problem is completely fixable simply by adding more power delivery circuitry. Give the RX 480 a second 6-pin connector and suitable corresponding VRMs and such on the board and you're set. It's not at all like the GeForce GTX 480 burning 300 W inside a radiator-like cooler that dared you to try frying an egg on it.
Back in the bad old days, video cards had fixed clock speeds that didn't adjust well for the particular workload, beyond clocking down at idle. Power viruses (e.g., FurMark, OCCT, or the StarCraft 2 title screen) that pushed a card harder than the company expected could fry things. But if you throttle clock speeds way back to handle the power viruses, you give up a bunch of gaming performance and people don't buy your cards.
Fortunately, AMD solved this in 2010 with PowerTune. It tracks power consumption in real time and throttles back clock speed by just enough to stay inside the desired power envelope. You don't get performance that obviously tanks as with the severe throttling from overheating. But you also don't need to know ahead of time everything that can push a card too hard. AMD demonstrably had it working way back in 2010 on the Radeon HD 6970.
If you set the PowerTune cap to 150 W, it shouldn't be possible for the card to pull 160 W for thermally significant periods of time. Did AMD drop PowerTune entirely? Is it malfunctioning? Did they increase the PowerTune cap to 160 or 170 W to try to score better reviews? Isn't it remarkable how these accidents tend to increase performance?
And now reports are coming in that the out-of-spec power draw from the RX 480 is damaging motherboards. It's probably only a tiny handful, and probably further restricted to cheap junk motherboards, likely backed by a mediocre or worse power supply, and possibly egged on by power weirdness coming from the wall. High quality components can handle running a little out of spec.
But you shouldn't rely on that. Even if you're going to overclock, you shouldn't run anything other than the component you're overclocking out of spec. If you want to overclock a CPU to the moon and have it burn 200 W, you should get a motherboard, power supply, case, and cooler that can handle a CPU putting out 200 W so that everything but the CPU itself is running in spec.
Running things out of spec unnecessarily is bad. Doing it intentionally on commercial hardware without telling anyone is worse.
Remember the GeForce GTX 590? It was a "365 W" card (already outside of the PCI Express specification, but not really bad in a desktop built around it) with two 8-pin PCI-E power connectors. That means that the rated power delivery was 375 W. That's quite a lot, but it didn't help that the 365 W TDP was a total lie and the card could easily blow well past 400 W. Some of them didn't survive the review process.
No, I didn't just say that Polaris is as bad as Fermi. But it isn't a good sign that that's the comparison I have to reach for.
Now, handling 400 W in a two-slot cooler is just plain hard. AMD finally got it right with the Radeon R9 295 X2 that liquid-cooled them both. But there's no excuse for not being able to handle 160 W.
So this is fixable with cards from board partners. And we should be thankful that AMD and Nvidia let partners such as MSI, Asus, Sapphire, and EVGA design and build cards. AMD and Nvidia don't always seem competent at it, and the reference cards mentioned above are far from the only clunkers in their histories. Remember the GeForce FX 5800 "dustbuster"?
But you know how else the running out of spec problem is fixable? Throttling back clock speeds more aggressively so that the card doesn't burn more than 150 W. That can be done with a driver update, and don't be surprised if AMD does exactly that.
The problem with cutting back clock speeds is that you lose performance. To stay inside of 150 W, maybe you lose 3% of your performance in this game and 5% in that one. And people notice lower numbers on bar graphs. If the performance losses don't come until after reviews are safely up and no one bothers to update them later, then they don't count, right? After all, the only people who suffer from that are your customers. They probably won't notice if they lose 3% of their performance, but they'll sure notice hardware failures.
So companies play various shenanigans to try to win reviews. Clock higher when you detect a canned benchmark running, and lower when you detect a power usage benchmark running. Both Nvidia and Intel have on various occasions said "look how fast it is" and "look how low power it is" for a part, trying to imply you could have both at once even though it wasn't even close to true. Make short-lived, small volume parts like the Radeon X800 XT PE. And launch just such a part as the claimed competitor to a competitor's real, volume part. Remember the EVGA GeForce GTX 460 FTW?
No, I didn't just say that Polaris is as bad as Fermi. But this is the third time I've had to assert that, and in comparison to four different cards from that architecture. The problems with the Radeon RX 480 are fixable by beefing up the power delivery, even without changing the clock speed. Third-party cards will do exactly that, and if history is any guide, probably at MSRP. Even if it adds $5 to the bill of materials, that shouldn't add $50 to the retail price tag. So I say, if you want a Radeon RX 480, you should wait for that. It probably won't be long.
1. Whatever happened to PowerTune?
2. Why does the reference RX 480 only have a single 6-pin PCI-E connector?
3. What if a driver update reduces clock speeds?
4. It's a good thing that third-party board partners are around.
Returning to the title, I'd include the GeForce GTX 1080, GeForce GTX 1070, and Radeon RX 480 in that. For the GeForce cards, it's really just a question of price. Do you really want to pay $700 (or $800 or $900, depending on how much you're gouged) for a product that you know will soon be $600?
But with the RX 480, it's something much worse. At least Nvidia was up front about pricing, if not timing. And the delays were incandescently obvious to those who understood the tech. But I've complained enough about Pascal, so I want to spend most of this post going after Polaris.
Some reviews noticed the Radeon RX 480 pulling 160 W. Now, there's nothing wrong with a desktop video card burning 160 W. But there's something very wrong with a PCI Express card with only a single 6-pin PCI-E power connector burning 160 W.
The PCI Express slot is rated as being able to deliver 75 W, a 6-pin connector also 75 W, and an 8-pin connector 150 W. If all you've got is the slot and a single 6-pin, that's 75+75 = 150 W. Pulling 160 W through that is running something out of spec.
Now, the Radeon RX 480 isn't the first card to do this. The GeForce GTX 470 had only two 6-pin PCI-E power connectors, and it routinely pulled more than 225 W.
No, I didn't just say that Polaris is as bad as Fermi. But I suspect that the reasons are the same: someone decided late in the game that the stock clock speed needed to increase. Rather than going back to the drawing board to beef up power delivery and make a card that could handle it, they just took the cards they had and clocked them higher.
And the problem is completely fixable simply by adding more power delivery circuitry. Give the RX 480 a second 6-pin connector and suitable corresponding VRMs and such on the board and you're set. It's not at all like the GeForce GTX 480 burning 300 W inside a radiator-like cooler that dared you to try frying an egg on it.
Back in the bad old days, video cards had fixed clock speeds that didn't adjust well for the particular workload, beyond clocking down at idle. Power viruses (e.g., FurMark, OCCT, or the StarCraft 2 title screen) that pushed a card harder than the company expected could fry things. But if you throttle clock speeds way back to handle the power viruses, you give up a bunch of gaming performance and people don't buy your cards.
Fortunately, AMD solved this in 2010 with PowerTune. It tracks power consumption in real time and throttles back clock speed by just enough to stay inside the desired power envelope. You don't get performance that obviously tanks as with the severe throttling from overheating. But you also don't need to know ahead of time everything that can push a card too hard. AMD demonstrably had it working way back in 2010 on the Radeon HD 6970.
If you set the PowerTune cap to 150 W, it shouldn't be possible for the card to pull 160 W for thermally significant periods of time. Did AMD drop PowerTune entirely? Is it malfunctioning? Did they increase the PowerTune cap to 160 or 170 W to try to score better reviews? Isn't it remarkable how these accidents tend to increase performance?
And now reports are coming in that the out-of-spec power draw from the RX 480 is damaging motherboards. It's probably only a tiny handful, and probably further restricted to cheap junk motherboards, likely backed by a mediocre or worse power supply, and possibly egged on by power weirdness coming from the wall. High quality components can handle running a little out of spec.
But you shouldn't rely on that. Even if you're going to overclock, you shouldn't run anything other than the component you're overclocking out of spec. If you want to overclock a CPU to the moon and have it burn 200 W, you should get a motherboard, power supply, case, and cooler that can handle a CPU putting out 200 W so that everything but the CPU itself is running in spec.
Running things out of spec unnecessarily is bad. Doing it intentionally on commercial hardware without telling anyone is worse.
Remember the GeForce GTX 590? It was a "365 W" card (already outside of the PCI Express specification, but not really bad in a desktop built around it) with two 8-pin PCI-E power connectors. That means that the rated power delivery was 375 W. That's quite a lot, but it didn't help that the 365 W TDP was a total lie and the card could easily blow well past 400 W. Some of them didn't survive the review process.
No, I didn't just say that Polaris is as bad as Fermi. But it isn't a good sign that that's the comparison I have to reach for.
Now, handling 400 W in a two-slot cooler is just plain hard. AMD finally got it right with the Radeon R9 295 X2 that liquid-cooled them both. But there's no excuse for not being able to handle 160 W.
So this is fixable with cards from board partners. And we should be thankful that AMD and Nvidia let partners such as MSI, Asus, Sapphire, and EVGA design and build cards. AMD and Nvidia don't always seem competent at it, and the reference cards mentioned above are far from the only clunkers in their histories. Remember the GeForce FX 5800 "dustbuster"?
But you know how else the running out of spec problem is fixable? Throttling back clock speeds more aggressively so that the card doesn't burn more than 150 W. That can be done with a driver update, and don't be surprised if AMD does exactly that.
The problem with cutting back clock speeds is that you lose performance. To stay inside of 150 W, maybe you lose 3% of your performance in this game and 5% in that one. And people notice lower numbers on bar graphs. If the performance losses don't come until after reviews are safely up and no one bothers to update them later, then they don't count, right? After all, the only people who suffer from that are your customers. They probably won't notice if they lose 3% of their performance, but they'll sure notice hardware failures.
So companies play various shenanigans to try to win reviews. Clock higher when you detect a canned benchmark running, and lower when you detect a power usage benchmark running. Both Nvidia and Intel have on various occasions said "look how fast it is" and "look how low power it is" for a part, trying to imply you could have both at once even though it wasn't even close to true. Make short-lived, small volume parts like the Radeon X800 XT PE. And launch just such a part as the claimed competitor to a competitor's real, volume part. Remember the EVGA GeForce GTX 460 FTW?
No, I didn't just say that Polaris is as bad as Fermi. But this is the third time I've had to assert that, and in comparison to four different cards from that architecture. The problems with the Radeon RX 480 are fixable by beefing up the power delivery, even without changing the clock speed. Third-party cards will do exactly that, and if history is any guide, probably at MSRP. Even if it adds $5 to the bill of materials, that shouldn't add $50 to the retail price tag. So I say, if you want a Radeon RX 480, you should wait for that. It probably won't be long.
Comments
Never thought of paying close attention to power pins like this.
Like i.e. Zotac is doing this all the time with their GTX series.
I've burned up a reference board because of power shenanigans, so I no longer buy them. I wait for manufacturers to come up with better cooling solutions, which seems to always fix the power problems.
The world is going to the dogs, which is just how I planned it!
https://youtu.be/qG2e-v94L4M?t=316
My hunch is that what i've read on a reddit post from some guy is what's happening here. Somehow the power draw programming stage got borked, and the card is going out of spec on the wrong part.
As in it was supposed to go out of spec and pull more on the 6-pin connector instead of the pci-e slot.
the 6-pin GPU connector cables for all but the worst psus are rated at 8Ax2x12v = 192W. Granted the psu itself might not be able to handle this, so the absolute maximum is a bit lower depending on the internal setup of a PSU.
What this tells you is that drawing 15,20,30w more on the 6-pin is not gonna make difference whatsoever ( the only difference between the 6-pin and the 8-pin is 2 ground wires, and them 2 wires don't miraculously transfer twice the amperage/wattage).
for instance, factory overclocked GTX750tis with no additional power connector draw more power from PCIe slot than rated 75W. Some 750ti have additional 6pin power connector, some dont and they all use same power, especially when overclocked.
Or that GTX950 with no additional power connectors which goes out of spec for PCIe slot and 75 W power draw on stock, but all reviewers commended it for good overclock when it pulls same power like every other OCed 950 that has 6 pin power connector, so 100-110W just from PCIe slot that is rated at 75W.
And using more power from PCIe cables is regular occurance on both sides. Probably why most stuff is overengineered and has high tolerance.
Generally, if something dies from this it was in some manner faulty in the first place, i mean, there was no mass frying boards from 750ti (and thats one of highest volume cards and one of reasons is because "it doesnt need additional power connector")
Reference cards, IMO, are not really good anyway, there are some arguments for mITX builds and SLI/CF since blower pushes hot air out of the case, but again, IMO, thats easily solvable by installing 2-3 fans in the case to make airflow sufficient for all your components.
Oh and if someone is wondering what the hell is the deal with a 6-pin vs 8-pin, if for a specific psu they used the same 192w rated wires, and are connected the same way inside the gpu to the 12v rail. Well the difference is that the middle wire in the 6-pin is not used.
If you used a 6+2-pin and tricked the gpu to start without the extra 2-pins, there would be no difference whatsover.
The same would happen if you would tweak the psu and the mobo and the gpu to use the middle wire in the 6-pin.
Or just go ghetto style and add 2 additional ground wires to trick the whole setup
i feel like i need to add that unless you really what you are doing, do not try to splice 2 more wires with the 6-pin cable :P
100/75=33,3% above the spec 100% of the time
things are worse with 950 with no power connector as OCed it can draw 100-110W on the same 75W PCIe slot.
Its really nothing new or extraordinary.
If you're pulling 160 W in total and the two places you can pull it from are only rated at 75 W each, you have to run something out of spec. It can be the slot or the 6-pin connector or both, but there's no way to do that and stay in spec. Only in the rarest of circumstances am I okay with silently going out of spec on power delivery.
If there is ample power delivery on the PCB, then they screwed up by not having a second 6-pin connector.
http://wccftech.com/amd-rx-480-pcie-power-issue-detailed-overclocking-investigated/
The GTX 750 Ti has a TDP of 60 W. That allows room for a considerable overclock without going over 75 W. And besides, it was a dumb card except as a test part for the Maxwell architecture.
The GTX 950 has a TDP of 90 W, and typically has a PCI-E power connector. A look through New Egg finds that most have a 6-pin connector, and many of the ones that don't have an 8-pin connector. Three of the other four have overclocks of at most 2 MHz (yes, two; not sure why they bothered) in gaming mode and 40 MHz in "OC mode", which is a bad idea to begin with when lacking a PCI-E connector, but might be okay if they're able to cherry-pick dies that they can undervolt far enough. Which they probably didn't.
And then there is this:
http://www.newegg.com/Product/Product.aspx?Item=N82E16814127937
Don't buy that card, and not just because it's overpriced.
The RX 480 isn't causing a mass frying of motherboards. But just because parts don't fry in the first few days doesn't mean it's okay. I care about reliability more than most consumers, so my take on this should be seen in that light.
Reference Sapphire 6870 lasted two years and 3 months. I wish I had taken a closer look at this card when I got it. The heat sink was not properly seated to the GPU and 3 of the 8 memory chips. 15 minutes fixing that may have gotten a couple of more years added to it's life.
XFX 7870 using their 'Ghost' cooling system and adding a 2nd 6 pin power plug is still going strong after 4 years... This card was absolutely put together properly.
The world is going to the dogs, which is just how I planned it!
https://www.techpowerup.com/reviews/ASUS/GTX_950/21.html
average 74W peak power in gaming 79W. But even that is out of spec of 75W on stock. If you go further in review:
https://www.techpowerup.com/reviews/ASUS/GTX_950/26.html
"24% overclcok on core and 25% overclock on memory"...on a card that goes out of spec on stock. On that OC it consumes power same as any other OCed 950 100-110W in gaming....all from PCIe slot. Noone thought its a big deal
I would say that this is most similar to 480. On stock it mostly stays below treshold but when OCed it goes way out of spec.
750ti....reference:
64W average and not a pretty graph that goes above the spec pretty much constantly. Now add factory OC and its pretty much way above spec. Now add manual OC and youre looking at a card that cunsumes 90-100 W. There are 750tis without aditional power connector and there are those which have it. 6pin power connector isnt there to look pretty, factory overclocked and manually overclocled 750tis use >> 75W
Moral of the story: if you want to overclock dont buy reference cards or cards that dont have power connectors (for which other in the class have like GTX950 and GTX750ti).
Boobs are LIFE, Boobs are LOVE, Boobs are JUSTICE, Boobs are mankind's HOPES and DREAMS. People who complain about boobs have lost their humanity.
The problem they are not specifying which 750 Ti those graphs are for, and they do mention the gigabyte one on the previous page.
And that gigabyte one has a additional 6-pin connector, which would make this graph irrelevant to say the least.
I mean the logic circuit which does the power redistribution may or may not be connected, and/or modifiable from bios, so technically flashing a new bios might work, but if there's no way to change it from the actual bios, only new cards from the factory will have this fixed, and considering this is reference, and that 75% of all planned reference cards are in assembly and finishing stage,
Don't Buy the Reference Card !!!
Oh but wait, there is way for the driver to fix this, downclock and downvolt the card enough so that overspec is 1-3w
So supposedly its GDDR5 vram that's the cause these power spikes, well at least partially responsible.
I'm pretty sceptical this is gonna solve the issue.