AMD Radeon RX Vega 56 and Vega 64 in the undervolting test

[Azix]

https://www.hardwareluxx.de/index.p...vega-56-und-vega-64-im-undervolting-test.html

Assessment of Undervolting
It is quite astonishing how much the gain by the undervoltings for the two cards fails. The Radeon RX Vega 64 and Vega 56 benefit greatly in terms of performance and savings in power requirements. It can be clearly seen that the Radeon RX Vega 64 is already quite close to the limit and there is little potential. In the Radeon RX Vega 56, on the other hand, we see a large scope, which should also be used by anyone who loves the building.

The results for the Radeon RX Vega 64 are as follows: In terms of performance, the GeForce GTX 1080 Founders Edition is quite clear. We have no values for an undervolting of the competition cards here, but take these values only as reference points. It would be unfair to compare a low-voltage card with one in the delivery state. At the same time, the undervolting of the Radeon RX Vega 64 means a significant reduction in power consumption, even if it is still part of the current single-GPU cards.

The Radeon RX Vega 56 is a little better. Here we see a significant reduction in the power requirement, which makes the GeForce GTX 1070 (again without undervolting) closer. The biggest jump, however, is at the same time as the performance, because despite a reduction in the voltage, we were able to significantly increase the cycle, or hold more stable at a high level. In the benchmarks, therefore, she moves the larger Radeon RX Vega 64 quite close to the Pelle and can also leave the GeForce GTX 1080 behind.

Currently, however, an undervolting is still quite complicated, since the software can not be trusted. Indicators for the clock do not have to be correct and if the voltage has been accepted, is only ensured by a glance at the consumption. A simple setting of the values in the software is not enough at the moment - everything has to be validated.

 

[thilanliyan]

Vega 56 is looking nice there. The 64 seems to be running near its lower voltage limit already.

 

[maddie]

https://www.hardwareluxx.de/index.p...vega-56-und-vega-64-im-undervolting-test.html

Higher perf/watt than the competition. Finding a way for low voltage operation appears to be the key. Have wondered if the auto voltage control mechanisms using IF are even working properly as yet, or are they borked. We'll see.

In contrast, Nvidia has this working properly for some time now.

 

[nathanddrews]

Wow, Vega56 is just the gift that keeps on giving. Uses only 5% more power than 1080Ti, but only 9-31% slower. Compared to the Vega64, which uses 39% more power and is 8-26% slower.

Too bad we'll never be able to buy one. I've added your link to the review mega-thread.

 

[nathanddrews]

So how is it that AMD's own engineers can't do this ahead of time via BIOS/drivers? Even with the silicon lottery, it seems to me that they could deliver much more efficient cards from the start. This has been true for many generations, but mostly starting with Hawaii. I don't buy it that it's just because Vega was rushed. It takes just as much time to undervolt/overclock as it does to overvolt/overclock. They ship the cards with dual BIOS switches and three primary power states in the driver... It just doesn't make sense, IMO.

 

[thilanliyan]

So how is it that AMD's own engineers can't do this ahead of time via BIOS/drivers? Even with the silicon lottery, it seems to me that they could deliver much more efficient cards from the start. This has been true for many generations, but mostly starting with Hawaii. I don't buy it that it's just because Vega was rushed. It takes just as much time to undervolt/overclock as it does to overvolt/overclock. They ship the cards with dual BIOS switches and three primary power states in the driver... It just doesn't make sense, IMO.

I'm guessing it's because of yields. Some of the cards probably need the stock voltage to hit the stock speeds, and to increase yields, they are also using the not so good dies.

 

[Phynaz]

So how is it that AMD's own engineers can't do this ahead of time via BIOS/drivers? Even with the silicon lottery, it seems to me that they could deliver much more efficient cards from the start. This has been true for many generations, but mostly starting with Hawaii. I don't buy it that it's just because Vega was rushed. It takes just as much time to undervolt/overclock as it does to overvolt/overclock. They ship the cards with dual BIOS switches and three primary power states in the driver... It just doesn't make sense, IMO.

Reliability. They have to make sure every chip functions as designed is whatever form the end user puts it in.

 

[Peicy]

Undervolting is great when it works for an individual card. AMD set stock voltages for a reason though, this does not make all Vega 56s more efficient in comparison to the 1070 since you can undervolt some of them too.

 

[Rifter]

I'm guessing it's because of yields. Some of the cards probably need the stock voltage to hit the stock speeds, and to increase yields, they are also using the not so good dies.

Exactly, you are going to need a golden sample to undervolt effectively. AMD set stock voltage for a reason, and that reason was to use the chips that binned so poorly they should never have been used, to build up stock to make this less of a paper launch.

 

[krumme]

Higher perf/watt than the competition. Finding a way for low voltage operation appears to be the key. Have wondered if the auto voltage control mechanisms using IF are even working properly as yet, or are they borked. We'll see.

In contrast, Nvidia has this working properly for some time now.

I think you perhaps are into something right here.
Notice they are applying 50% perf in the settings.

At standard settings read in wattman my rx64 runs at aprox 970Mhz 98% of the time a few moments 1200. Is the freq reading borked in Wattman? - i dont think so. The reason:
When i apply +50% in wattman - all else equal - and make sure temp is set high and fan high- i get a constant 1632 MHz. I can alter voltage to 1050 from 1200. Makes no difference.
Needless to say the difference for gaming performance in bf1 is littarally going from low to ultra at 60 fps. at 4k. (dx11 MP64 maps)

 

[tential]

I'm guessing it's because of yields. Some of the cards probably need the stock voltage to hit the stock speeds, and to increase yields, they are also using the not so good dies.

They're also just bad at this. At this point with this many years of experience, we know this to be the case. We were relatively certain the card was going to be poorly managed power wise before the card even came out.

Amd makes poor decisions sometimes that can only be explained by severe incompetence

 

[tential]

I think you perhaps are into something right here.
Notice they are applying 50% perf in the settings.

At standard settings read in wattman my rx64 runs at aprox 970Mhz 98% of the time a few moments 1200. Is the freq reading borked in Wattman? - i dont think so. The reason:
When i apply +50% in wattman - all else equal - and make sure temp is set high and fan high- i get a constant 1632 MHz. I can alter voltage to 1050 from 1200. Makes no difference.
Needless to say the difference for gaming performance in bf1 is littarally going from low to ultra at 60 fps. at 4k. (dx11 MP64 maps)

I thought it was common knowledge to always run the power limit high for amd cards. It's limited downside great upside for clock stability and overclocking...

 

[Kallogan]

really nice vega 56 not only undervolting seems to up performance close to a 1080 but the power draw is close to 1070.

 

[krumme]

I thought it was common knowledge to always run the power limit high for amd cards. It's limited downside great upside for clock stability and overclocking...

I can say it another way. There is a possibility all reviews that didnt appply +50% power limit of the rx64 (and who did) had this card run at mostly 970 max 1200 a few times.

Mine did that for standard setting at least in bf dx11 mp64 - 970 all the time. And i am prettry sure about that. When i apply 50% i get a perf that is way over what eg AT did. Way above. But wattman also say 1632 consistently. Surely this needs more testing.

I say it because look at those voltage numbers in the op link. They give no sense imo. Had they run with standard voltage they would probably have gotten better results eg the ex64 from the 1570 or so to 1632.

Their performance results is from adding the headroom. Things gets mixed here.

 

[krumme]

When gamersnexus did their underclock of the rx56 they applied +50% too btw.
http://www.gamersnexus.net/hwreviews/3020-amd-rx-vega-56-review-undervoltage-hbm-vs-core

 

[vinhom]

Isn't this an effect of AVFS being disabled?

 

[Yakk]

really nice vega 56 not only undervolting seems to up performance close to a 1080 but the power draw is close to 1070.

After the huge disappointment that was Vega 64, this gives me hope. If the price is really what AMD said... and it is available to buy...

 

[JDG1980]

RTG does this over and over again.There seems to be a serious cultural issue there; apparently no one there has gotten the memo that perf/watt matters, and that binning exists. They insist on scraping the bottom of the barrel and using every bit of trash silicon, instead of throwing away the worst ~10% (or at least reserving them for a low-cost SKU) to get much better performance on the rest. It's just RTG that does this; the CPU division of AMD seems to have done a fine job with Ryzen. Sure, 1800X is clocked past its perf/watt sweet spot, but 1700 is just about perfect - and none of these chips have far more voltage crammed through them than they really need. RTG has been talking perf/watt for a while, and Polaris was supposed to bring some sort of auto-voltage calibration, but it's all been hot air - they're just shoving a ton of extra juice through every gaming card, no matter what.

 

[Olikan]

that binning exists.

Err... the good ones are probably going to Apple

 

[JDG1980]

Err... the good ones are probably going to Apple

Not all of them - it's been demonstrated that regular, off-the-shelf Vega cards can undervolt well. This is unsurprising in the case of Vega 56, which has lower stock clocks than Vega 64.

It's not that only "golden" chips can undervolt. Rather, all but a tiny handful of trash chips can undervolt, and those trash chips should have been discarded instead of allowing them to drag down everything else.

 

[maddie]

I think you perhaps are into something right here.
Notice they are applying 50% perf in the settings.

At standard settings read in wattman my rx64 runs at aprox 970Mhz 98% of the time a few moments 1200. Is the freq reading borked in Wattman? - i dont think so. The reason:
When i apply +50% in wattman - all else equal - and make sure temp is set high and fan high- i get a constant 1632 MHz. I can alter voltage to 1050 from 1200. Makes no difference.
Needless to say the difference for gaming performance in bf1 is littarally going from low to ultra at 60 fps. at 4k. (dx11 MP64 maps)

Didn't bother saying anything earlier as too many shouting down. I get the impression that internal settings are conflicting.

Isn't this an effect of AVFS being disabled?

My thoughts also, or at least screwed-up. If it's just the bios/driver that needs to be sorted, then we have a lot of performance to be exposed. If it's the actual control circuitry, then no hope. The fact that it can be done manually allows some hope for a fix as the basic hardware appears capable of a lot more than we've seen.

 

[Magic Hate Ball]

Not all of them - it's been demonstrated that regular, off-the-shelf Vega cards can undervolt well. This is unsurprising in the case of Vega 56, which has lower stock clocks than Vega 64.

It's not that only "golden" chips can undervolt. Rather, all but a tiny handful of trash chips can undervolt, and those trash chips should have been discarded instead of allowing them to drag down everything else.

Well, a large number of RX460's could be fully unlocked, despite the batches of fully operational ones going to Apple.

I think once they reach their quota of binned ones, they just don't have enough left over to financially justify tiering. Like when Nvidia did the GTX 560 ti 448 Core Edition (I had it) that was really a slightly more cut down 570/580 (GF110) dies they had lying around.

Depending on yields, I wonder if a RX Vega 48 w/ 4gb of HBM2 could fill in just above the RX580.... maybe once they figure out HBCC better.

 

[beginner99]

Gosh. How can a company constantly fail this bad at releases and PR? There is only 1 explanation for this: hardwareluxx had a sub 1% Golden Chip. This test needs to be repeated with different cards.

Radeon RX Vega 56 P6: 1.537 MHz
P7: 1.592 MHz
P6: 1.150 mV
P7: 1.200 mV 394,3 W

Radeon VX Vega 56 Undervolting P5: 1.613 MHz
P6: 1.613 MHz P5: 1.070 mV
P6: 1.070 mV
-73 W

They were able to reduce voltage by 0.13 (from 1.2 mv to 1.07mv) while at the same time increasing clocks to 1613 mhz from 1592 mhz (ok only 1.3%). With that they save 73W (18%). It beats a 1080 founders clearly while using less power than it. Also it perfroms a lot better than stock Vega 56. I think the reduced power actually makes it run it higher clocks constantly. It gets a >10% boost in all games, some near 20% with 1.3% higher clocks (or 4.9% higher to base clock). Their undervolting clearly reduces throttling.

There really can be only 2 reasonable explanation for this: Golden Chip or utter stupidity throughout RTG from Tech to PR.

 

[maddie]

Gosh. How can a company constantly fail this bad at releases and PR? There is only 1 explanation for this: hardwareluxx had a sub 1% Golden Chip. This test needs to be repeated with different cards.

Radeon RX Vega 56 P6: 1.537 MHz
P7: 1.592 MHz
P6: 1.150 mV
P7: 1.200 mV 394,3 W

Radeon VX Vega 56 Undervolting P5: 1.613 MHz
P6: 1.613 MHz P5: 1.070 mV
P6: 1.070 mV
-73 W

They were able to reduce voltage by 0.13 (from 1.2 mv to 1.07mv) while at the same time increasing clocks to 1613 mhz from 1592 mhz (ok only 1.3%). With that they save 73W (18%). It beats a 1080 founders clearly while using less power than it. Also it perfroms a lot better than stock Vega 56. I think the reduced power actually makes it run it higher clocks constantly. It gets a >10% boost in all games, some near 20% with 1.3% higher clocks (or 4.9% higher to base clock). Their undervolting clearly reduces throttling.

There really can be only 2 reasonable explanation for this: Golden Chip or utter stupidity throughout RTG from Tech to PR.

That threw me for a second. For the card it's obviously much more than 18%. You're using system power to calculate that %.

 

[krumme]

My thoughts also, or at least screwed-up. If it's just the bios/driver that needs to be sorted, then we have a lot of performance to be exposed. If it's the actual control circuitry, then no hope. The fact that it can be done manually allows some hope for a fix as the basic hardware appears capable of a lot more than we've seen.

I can see i get thermally limited at the standard settings. If i manually up fanspeed 40% and set temp 5-10c higher, i can get it from 970/1200 to 1530 while playing. Still not 1632MHz it seems. I need >=+25% to get there.

Perhaps we have a mix of problems here adding up; an avfs not working quite as it should combined with amd trying to get so high freq as possible and high yield as possible by upping voltage very high. Its comical to think of. AMD add voltage to get from eg 1570 to 1632 to get freq and also yield (actually i think the voltage is set for 1677 for the liquid version), and then you have a agressive avfs system that throtttles back. So you end with a slower card that uses more power.