AMD's New Radeon RX 480 Driver Fixes Power Issues

[TJ Hooker]

The idea behind the standards of the PCI SIG is, that the connector isn't delivering pure DC, but a kind of HF current with extreme and fast load level changes (up to 100 KHz). We also tested it internaly with a good lab PSU and resistors - no problem. But the temperatures are going up, if you are using pulsing current, especially on older, used plugs and sockets.

Could you elaborate on this? When you say pulsing current, do you mean you applied a pulse train current waveform to the resistor? And when you say the temperatures are going up, going up compared to what?

 

[FormatC]

I'm not a native speaker, sorry. To geht the same HF current and not only flat DC we bundled the two 8-pins of a R9 295 X2 with the help of two cables to a dummy PCB with a single 6-pin connector. The PCIe cable of an older single-rail Corsair AX1200 (Flextronics) can deliver each current you need, in this case up to 450 watts (or approx 38 A). This was one of the experiments we does with the R9 295 X2, may be you remember my video with a 500 watts PSU and this VGA beast.

The temperatures were monitored with IR equipment (cable, socket, plug). If you are using AWG16 cables the botleneck is the socket, not the cable itself or the the pins of the plug. The resistance of this connection increases with the frequency. If you are using absolutely new sockets and plugs it may work better, but for my taste are above 110°C simply too much.

 

[Sakkura]

And BTW, the 8-pin allows you max. 150 Watts due two additional ground wires.

The official PCI-SIG spec says that. The Molex spec on the other hand says that the 6-pin connector is fine up to 9A per standard pin and 13A for high-current variants, which are about 300W and 450W respectively.

The PCIe spec is being extremely conservative by specifying a current that's only 1/4 of what the connector's lower-end spec is actually capable of. The two extra grounds on the 8-pins connector don't really add much in terms of useful current-carrying capacity since up to two of them may be wasted on connector identification anyway.

The idea behind the standards of the PCI SIG is, that the connector isn't delivering pure DC, but a kind of HF current with extreme and fast load level changes (up to 100 KHz). We also tested it internaly with a good lab PSU and resistors - no problem. But the temperatures are going up, if you are using pulsing current, especially on older, used plugs and sockets. The second fact is, that nobody knows, which cable quality is used on the PSU side. AWG 16, 18 or 20? In the case of AWG 20 and two wires I wish you a lot of fun and a good weather if you make the same test

The discussion about industrial standards us useless. And AMD has never requested a PCI logo and the PCIE SIG also said, that this card will not pass the certification with this current specs.

AWG 20 goes against recommendation and is something that gets exposed in PSU reviews.

 

[InvalidError]

The idea behind the standards of the PCI SIG is, that the connector isn't delivering pure DC, but a kind of HF current with extreme and fast load level changes (up to 100 KHz).

If you have high current going through loose wiring at 100kHz, you have a massive EMI compliance issue in the making, not to mention the voltage drop from having hundreds of nanohenries worth of parasitic inductance. The wires are useless for high speed transients, that's why sufficient bypassing to absorb them is necessary at both ends.

Another detail to keep in mind: when measuring current with a current transformer, you add parasitic inductance to the wires you are measuring and this will alter the resonant circuit formed by the capacitors at both ends of the wires, the wires in the loop and the current transformer. This high-Q resonant circuit may cause high current ringing that wouldn't be there or would be of much lesser magnitude under normal operating conditions.

 

[FormatC]

AWG 20 goes against recommendation and is something that gets exposed in PSU reviews.

Tell this the stupid idiots in the SI business. Low-cost on each edge. The end-user market is really small in comparison to this PC assemblers. Take a look in such tin cans and you will be shocked But exactly these low-cost manufacturers are the best customers for the 480 reference.

@InvalidError:
The VR is switching up to 200 KHz. Ok, a lot is filtered, but it is a fact that a VGA card produces a very fast switching current. I was in FSP groups R&D and safety lab two years ago and we measured this currents (because they proofed my measurings, seen in our reviews). Alls this data were the reason to change the PSU design and to change also the models of their secondary caps.

 

[virajk]

Can we expect non-reference boards to solve the issue ? Will the future reference boards solve the issue as well ?

 

[FormatC]

I hope that AMDs partners will do it right (and better). But we have no info about the availability of these cards. We were not able to order one sample, nothing. And a lot of companies are not focussed on the RX 480. That means: no media samples yet from Gigabyte, MSI and others.

 

[rnewton8]

I dont know where all these people are getting that enabling compatibility mode with RX 480 only yields a small 3% performance hit. I have a rx 480 and enabling compatibility mode causes a massive performance hit. I went from 10k in firestrike to 7k, so thats closer to 30%, not 3%. With comp mode turned off and new driver, performance is the same

 

[FormatC]

Firestrike isn't a game, it is a "power virus". As higher the permanent load as lower the performance. In a normal game you have a lot of load level changes and you will not mention such a big difference

 

[tsnor]

I dont know where all these people are getting that enabling compatibility mode with RX 480 only yields a small 3% performance hit. I have a rx 480 and enabling compatibility mode causes a massive performance hit. I went from 10k in firestrike to 7k, so thats closer to 30%, not 3%. With comp mode turned off and new driver, performance is the same

Wow, that's a hit. Rough rule of thumb, doubling the power gives 30% performance, so your performance hit of 30% is comparable to cutting the power load to the card in half. Very strange.

10K is what other pre-compatibility mode reviews were seeing, so your results are consistent that way. Interesting.

 

[mluppov]

For some reason not a single word about overclocking, which is obviously not an option here. Actually we know the reason why they did not mention overclocking, do we? "Sponsored" by AMD. The card is a total failure. At the very least they could have acknowledge this and offer refund/replacement. But no, they chose to act pretty much like nvidia does by pretending nothing really happened. I wonder if they continue selling these messed up cards. In any case now when GTX 1060 came out, only a total fool will buy RX 480.