AMD RX 480 Fails PCI-E Specification

[Angilion]

If you've got a modern motherboard you should be okay; iirc GPUs can legitimately pull 150w from PCIe 2.1 and above. That having been said If Tom's is accurate it's a bit of a screw up on AMD's part still.

Can anyone provide a link to any official statement from PCI-SIG about that?

The bits I can find state a 300W limit for total power draw. Not from the PCI-E slot. 300W card supply is given as 75W from slot + 75W from 2x3 pin + 150W from 2x4 pin (preferred) or 75W from slot + 75W from 2x3 pin + 75W from 2nd 2x3 pin + 75W from 3rd 2x3 pin (not preferred).

 

[triss]

Can anyone provide a link to any official statement from PCI-SIG about that?

The bits I can find state a 300W limit for total power draw. Not from the PCI-E slot. 300W card supply is given as 75W from slot + 75W from 2x3 pin + 150W from 2x4 pin (preferred) or 75W from slot + 75W from 2x3 pin + 75W from 2nd 2x3 pin + 75W from 3rd 2x3 pin (not preferred).

From the video where they showed all the wattage etc iirc the only statement that was found was the pin connector specs itself .

 

[Angilion]

You can never defeat Nvidia fans with arguments. They will use their own illiteracy to come at you in great numbers and scream at you that you should have bought none other than Nvidia card.

I have an AMD card (7950).

The card I had before this one was an AMD card (4870).

The card I had before that was an AMD card (9800 Pro).

The card I had before that was an AMD card (7900 GT).

The card I had before that was an AMD card (7800 GT).


You are talking nonsense. You are behaving in exactly the way you are complaining about and projecting your own behaviour onto everyone else. You are wrong.

I bet there will not be apologies or retractions from all the websites that published this nonsense, when from the first sight it is clear that PCIE allows far more than 75w from socket more like 300w.

You haven't even looked at the PCI-E spec, have you? Your "first sight" consists of seeing what you want to exist, not what actually exists.

 

[sqwerty]

Can anyone provide a link to any official statement from PCI-SIG about that?

The bits I can find state a 300W limit for total power draw. Not from the PCI-E slot. 300W card supply is given as 75W from slot + 75W from 2x3 pin + 150W from 2x4 pin (preferred) or 75W from slot + 75W from 2x3 pin + 75W from 2nd 2x3 pin + 75W from 3rd 2x3 pin (not preferred).

Can't get anymore official than the spec itself.

Power for cards that support a 75W maximum power dissipation can be drawn via a combination of +12V and +3.3V rails but each rail draw is limited and the sum of the draw on the two rails cannot exceed 75W.

Not sure however if they have revised that figure since 2004? as we've had numerous revisions of the overall PCI-E spec since then, but I doubt there'd be this much fuss if they had upped that figure... right?

 

[Rroff]

Can't get anymore official than the spec itself.

Power for cards that support a 75W maximum power dissipation can be drawn via a combination of +12V and +3.3V rails but each rail draw is limited and the sum of the draw on the two rails cannot exceed 75W.

Not sure however if they have revised that figure since 2004? as we've had numerous revisions of the overall PCI-E spec since then, but I doubt there'd be this much fuss if they had upped that figure... right?

End of the day you have to consider the weakest link even if some systems are compatible with a revised spec.

 

[jbh110tdi]

I looked at the RX480 playing on the idea on getting one for my lad to upgrade his 7950.
His board is a few years old ( asus crosshair v formula z )

Im glad I held off and this has put me off the RX480 so now ill go back to my original plan of saving up for a gtx 1070 and hell get my 970

 

[mark42]

Can anyone provide a link to any official statement from PCI-SIG about that?

The bits I can find state a 300W limit for total power draw. Not from the PCI-E slot. 300W card supply is given as 75W from slot + 75W from 2x3 pin + 150W from 2x4 pin (preferred) or 75W from slot + 75W from 2x3 pin + 75W from 2nd 2x3 pin + 75W from 3rd 2x3 pin (not preferred).

The relevant spec: "PCI Express x16 Graphics 150W-ATX Specification"
(it's not freely available)

for the PCIe connector: 3A from 3.3V and 5.5A from 12V
6-pin connector: 6.25A from 12V

8% voltage sag is permitted, so worst case total power available from 12V is:
(5.5A + 6.25A) * 12V / 1.08 = 130W

You can get up to 10W from 3.3V.

 

[Angilion]

Im interested in how much these might overclock now when they get another power connector and water cooling. They might turn out to be a 300watt overclocking beast.

They're power throttled at stock as well, so merely adding another power connector and increasing the allowed power draw will improve performance without any overclocking.

I wouldn't buy a reference RX 480, but I am eager to see what's what with a 3rd party RX 480 that has the power supply that an RX 480 should have. Simply replacing the 6-pin connector with an 8-pin one or adding a second 6-pin one would do. I don't care if it draws an extra 30W, which is about as much extra as it will be allowed to draw with the power limit set to +20%. 30W. Whoopie. I'm not running it off batteries. I don't care if it's a 180-190W card rather than a 150-160W card.

You'd need some modification to get it to draw 300W, either in the BIOS or through an overclocking tool that allows you to specify a power limit increase of more than 20%. Doable, though.

Personally, I'd be looking at removing the power throttling first by a combination of undervolting and an increased power limit, while keeping stock speeds. Then I'd edge up clocks and voltages while monitoring temps and power throttling. There's not much point increasing clocks if doing so increases power throttling - you get lower performance and higher temps.

Just using an 8-pin connector (or 2 6-pin connectors) would have kept the power draw through the PCI-E slot below 75W and avoided this whole issue. But no, marketing required AMD to use a single 6-pin slot to "prove" that the card only uses a maximum of 150W even though it needs to have its performance cut down to do so and it doesn't always work properly anyway. A reference RX480 is either more severely reduced by power throttling (to max out at 150W) or it exceeds AMD's claims and the PCI-E spec and is still reduced by power throttling (to max out at ~165W). Daft. Should have just called it a 200W card and sold it on price/performance. Without the power throttling imposed to create an artifically reduced power draw, the card's performance would be increased but the price wouldn't, so it would be pretty compelling on that basis.

 

[Angilion]

The relevant spec: "PCI Express x16 Graphics 150W-ATX Specification"
(it's not freely available)

for the PCIe connector: 3A from 3.3V and 5.5A from 12V
6-pin connector: 6.25A from 12V

8% voltage sag is permitted, so worst case total power available from 12V is:
(5.5A + 6.25A) * 12V / 1.08 = 130W

You can get up to 10W from 3.3V.

I'm not aware of a seperate 6-pin connector that's part of the PCI-E slot on the motherboard. Am I missing something or are you mistakeningly quoting the specification for total power draw from all sources rather than power draw through the PCI-E slot?

If it's the latter then the figures in the spec you quote shows a limit of 75.9W for power drawn from the PCI-E slot: (5.25 * 12) + (3 * 3.3) = 75.9

 

[Angilion]

Your motherboard will be absolutely fine. The M5A97 R2.0 has PCI-E 2.0 slots, which can supply up to 300W perfectly happily (last paragraph here: http://www.tomshardware.com/reviews/pci-express-2.0,1915-2.html). This would only be a potential issue if you were still running a really old motherboard with a PCI-E 1.0 slot.

The paragraph you refer to directly contradicts you. The PCI-E specification also directly contradicts you.

I think it's fair to say that your statement, above, is simply wrong.

That last paragraph refers to the total power draw, which is why it clearly states "Graphics board designers still have to work with auxiliary power connectors". PCI-E 1 was 75W from slot + 75W from 6-pin connector. PCI-E 2 added the 8-pin connector for another 150W. The PCI-E 2 spec allows for 300W graphics cards, but 225W of that comes from seperate connectors directly from the PSU. It's still 75W from the slot. Your statement is wrong.

 

[Angilion]

The fact is that the RX 480 pass the PCIe compliance test.

Not if it's tested for compliance with the power draw through the PCI-E slot. So presumably it wasn't. Or an engineering sample was and that sample differed from the retail cards. All that would be required would be to use a different BIOS that power throttles the card even more. You could get the power draw down to almost anything you want that way.

Oddly, it seems that PCI-SIG aren't very concerned about compliance with the power draw limits in the PCI-E standards anyway.

 

[mark42]

I'm not aware of a seperate 6-pin connector that's part of the PCI-E slot on the motherboard. Am I missing something or are you mistakeningly quoting the specification for total power draw from all sources rather than power draw through the PCI-E slot?

If it's the latter then the figures in the spec you quote shows a limit of 75.9W for power drawn from the PCI-E slot: (5.25 * 12) + (3 * 3.3) = 75.9

I thought it would be obvious that the 6-pin connector is the one from the power supply and not part of the slot.

To be exact, from the slot only you get:
12V * 5.5A / 1.08 + 3.3V * 3A / 1.09 = 70.2W
with worst case voltage sag (the 3.3V volt rail is 9% vs. the 8% for 12V.)

The specification limit is a current limit and only indirectly a power limit.

 

[Angilion]

Is this likely to be present on aib cards?

A card with an 8-pin connector (or 2 6-pin connectors) and configured to draw a higher proportion from that connector (or connectors) would be fine.

The bit in bold might be a problem. Or it might not. I don't know how the power draw is determined, so I don't know if it could be different on a different card. The reference card is getting ~50% of its power through the PCI-E slot and if that proportion isn't reduced then the problem will remain. If it is reduced and the the right connector(s) are on the card, the problem will disappear. 225W with 75W from the slot and 150W from the connector(s) would be fine. 225W with 112.5W from the slot and 112.5W from the connector(s) would not be fine (although it might work if the motherboard was well over spec).

 

[Angilion]

great, thanks for that and good to see it's a software fix. I really hope they manage to get rid of this issue.

Don't be too sure about counting your chickens yet.

There are two possible approaches that don't require new hardware:

1) Power throttle the cards more than they already are. This will limit maximum power use to below 150W, making it possible to keep power draw within spec for both the PCI-E slot (75W) and the 6-pin connector (75W). Of course, this approach reduces performance.

2) Alter the proportion of the total power draw taken from each source. So instead of drawing 85W from the PCI-E slot and 85W from the 6-pin connector, draw 75W from the PCI-E slot and 95W from the 6-pin connector. This approach goes over spec for the 6-pin connector even more than it already is. That probably won't matter because they're usually capable of going way over spec without any issues. But they're not guaranteed to do so because they're only required to be in spec, so it may cause problems.

EDIT: Bugger, I am wrong and missed the obvious fix that avoids both problems - reduce the voltages if that's possible. Wouldn't take much of a reduction.

 

[Angilion]

The fix should be quite easy for them. People on Reddit have tested undervolting the card and have actually gained performance while lowering the power usage, temps and fan speed.

The best result I've had personally was a 36% increase in performance, a 12C reduction in GPU delta and a 19C reduction in VRM delta. That was on a 7950.

Power throttling is an extremely imprecise blunt instrument of a way to reduce power consumption. Reducing the amount of power actually required (by undervolting) is a far better way of reducing power consumption. It requires experimentation and testing for each individual card, so manufacturers are stuck with using high voltages and really heavy-handed power throttling to make sure that every card will run at the stated clocks. That 7950 of mine is clearly towards the upper end of the curve in terms of how low it can go on volts for a given clock speed (hence the size of the improvement) but you should see some improvement on any card that power throttles. Undervolting is the new overclocking

 

[Angilion]

What do you think about all the speculation on why the card behaves how it does? intrinsic design error, or late-stage increases in voltage to ensure performance was in line with the marketing (to a point the card wasn't designed for)

I think it's a bit more messy than that because marketing was also based on power usage and I think that turned out to be higher than AMD expected. So they had to try to squeeze out enough performance to justify the hype while squashing power consumption down to the claimed levels and to do both from a card that used more power than it was expected to use, all the while piling a high enough margin of error on top of the voltage requirements to ensure that every card was completely stable.

From any point of view other than marketing, the reference card has an intrinsic design error - it doesn't have the right power connector, so it's not capable of supplying enough power to its components while remaining in spec even when it's being throttled to reduce power draw, let alone when it's at it's full potential.

 

[Meaker]

They can't balance the power draw on this card via software. The pci-e and 6 pin are not physically linked.

The pci-e slot powers the AUX rail and VRAM so would need some kind of change in memory power limit if such a thing exists.

With the current design you could have 2x 8 pin and if would still draw too much from the pci-e. You need to fill the bridging resistors near the pci-e plug to modify the balance. However then the card will attempt to run at full speed with the 6 pin not plugged in as it has no sense pin (it uses it for extra power draw and turns the 6 pin like an 8 pin with e conducting pairs).

 

[FredFlint]

I don’t know if they will be able to fix the problem properly. A fix would be to re-balance the power so it works like all the other GPU's and uses < 66W from the PCI-E port and gets the rest from the 6-pin. I will wait and see what the new fix is and then decide if I should return my card.

 

[Meaker]

Nevermind.

 

[Noughtboy]

That is really useful. A picture paints a thousand words and all