AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

This is weird, from the texturing, the tesselation, the bw,... I hope we will have great reviews explaining all this (exept if RX Vega correct this but I don't see how..)
 
only a few more days. with the vega tour AMD is doing, maybe we get something before 30th. I am expecting at least significantly better performance from rx vega. too many weird results from vega fe.

not sure if posted here but on http://www.pcgameshardware.de/Vega-...ase-AMD-Radeon-Frontier-Edition-1232684/2/#a3

Update 13.07.2017: Undervolting

Meanwhile, we have been dealing with the Undervolting of the Frontier Edition for a few hours and can tell a lot of good things. After the card had already completed all 1.600 MHz game tests at only 1.1 instead of 1.2 volts (see benchmark section of this article), we have further reduced the voltage test-by-step. With 1,075 Volt via Wattman, our retail store is still rockstable, all game benchmarks run smoothly and without graphics errors at 1,600 MHz. The power consumption of the entire PC is at the same level as the factory setting, although we have set the power limit to +50 percent. The latter ensures that the Radeon Vega Frontier Edition does not slow down temporarily, it should run into the power limit. We also allow the card to have a higher fan speed. With 1.060 Millivolt (whether this voltage is set 1: 1 or directly to 1.050 Millivolt, we do not know yet) came to the first crash.

The results show, firstly, that the chip is enormously profit from undervolting and secondly that Vega can be significantly higher than Polaris - the latter is hardly stable above 1.45 GHz to get, let alone to 1.6 GHz with so little voltage. We will further examine the power consumption and efficiency gains.
Click to expand...

they managed to get it running 1600Mhz with 1075 mV. Already we have some range for AMD to play with. Apparently it couldn't do compute at that voltage without issues but runs games just fine with it. Below is the normal behavior they observed before undervolting.

1.600 MHz @ 1.200 mV: Bestcase-Boost, is used under load only briefly / rarely
1.528 MHz @ 1.150 mV: Next lower stage, is usually used after a few seconds
1.440 MHz @ 1,100 mV: Boost level very frequently encountered under full load
1.348 MHz @ 1.050 mV: under full load comparable to 1.440 MHz
1,269 MHz @ 1,000 mV: Lowest of us recorded boost stage under load, only rarely used
1.137 MHz @ 950 mV: Second-to-last boost stage, never used in the test
992 MHz @ 900 mV: Lowest boost level, never used in the test
Click to expand...

Should be really interesting when details come out. AMD has turned up the suspense on this one. Full blown mystery.
 
Last edited:
Google translated from german:
Meanwhile, we have been dealing with the Undervolting of the Frontier Edition for a few hours and can tell a lot of good things. After the card had already completed all 1.600 MHz game tests at only 1.1 instead of 1.2 volts (see benchmark section of this article), we have further reduced the voltage test-by-step. With 1,075 Volt via Wattman, our retail store is still rockstable, all game benchmarks run smoothly and without graphics errors at 1,600 MHz. The power consumption of the entire PC is at the same level as the factory setting, although we have set the power limit to +50 percent. The latter ensures that the Radeon Vega Frontier Edition does not slow down temporarily, it should run into the power limit. We also allow the card to have a higher fan speed. With 1.060 Millivolt (whether this voltage is set 1: 1 or directly to 1.050 Millivolt, we do not know yet) came to the first crash.

The results show, firstly, that the chip is enormously profit from undervolting and secondly that Vega can be significantly higher than Polaris - the latter is hardly stable above 1.45 GHz to get, let alone to 1.6 GHz with so little voltage. We will further examine the power consumption and efficiency gains.
Click to expand...

I thought no one was being able to do vcore changes on Vega FE cards?
Stable 1600MHz with 1.075V sounds good. I hope they release power consumption numbers for these settings but without increasing the power limit.
 
Stable 1600MHz at 1.075V (which resulted in 250W in their testing )would at least not be a regression from Polaris in perf/watt, I believe? It's sad that that's even a thing lol

I also doubt that tour will include any proper performance comparisons. I.E FPS counters and A/B testing systems. It will probably just be a bunch of machines with pre-selected games without any performance metrics, and a camera shooting a bunch of players playing said games.

That would imply a far more confident AMD than what we're seeing with Vega so far.
 
Last edited:
Cat Merc said:
Stable 1600MHz at 1.075V (which resulted in 250W in their testing )would at least not be a regression from Polaris in perf/watt, I believe? It's sad that that's even a thing lol

I also doubt that tour will include any proper performance comparisons. I.E FPS counters and A/B testing systems. It will probably just be a bunch of machines with pre-selected games without any performance metrics, and a camera shooting a bunch of players playing said games.

That would imply a far more confident AMD than what we're seeing with Vega so far.
Click to expand...
They will only show real perf. comparisons if Vega magically manage to beat or be close the the 1080Ti. which I doubt. I'm quite sure they will play the 4k max detail at 60hz lock thing once more.
 
mczak said:
I don't see anything in the test indicating they even tried to run compute benchmarks with undervolting...
Click to expand...

Malo said:
Where are you seeing this? Maybe it's a translation issue but I don't see this mentioned in the article.
Click to expand...

its from the forum the author posts in. "Raff" is the one linked to the pcgameshardware.de site

https://www.forum-3dcenter.org/vbulletin/showthread.php?p=11429132#post11429132

there are further discussions beyond that post.
 
The same happened with Polaris, AMD put too much voltages and most of the cards(early cards) actually gained performance when undervolting. But looking at vega it seems that they really rush the launch(again).
 
xEx said:
The same happened with Polaris, AMD put too much voltages and most of the cards(early cards) actually gained performance when undervolting. But looking at vega it seems that they really rush the launch(again).
Click to expand...
The thing is - most only tested that with gaming, we don't know if the card is stable on other loads (compute) with those lower volts and AMD needs to make sure it is
 
Kaotik said:
The thing is - most only tested that with gaming, we don't know if the card is stable on other loads (compute) with those lower volts and AMD needs to make sure it is
Click to expand...
Well the card was not fully stable with launch drivers either.
 
Kaotik said:
The thing is - most only tested that with gaming, we don't know if the card is stable on other loads (compute) with those lower volts and AMD needs to make sure it is
Click to expand...
Wasn't that what AVFS was designed to do without the heavy handed approach of forcing clocks and voltages? One other possibility with the overbuilt power phases is that each cluster has an independent supply. There could be a lot of voltages capable of being adjusted. With XB1X and the Hovis design they separated CPU and GPU. With Infinity they could take it a step further.
 
Raff (he is my colleague) has posted in the 3DCenter thread, that at first, in fact, the card hard crashed (i think he meant blackscreen) with 1,075 mVolt AND at +50% power limit in the ComputeMark subtest MandelS Complex - when run in benchmark mode. He then tried the stress testing mode, and there it ran fine even down to 1,050 mVolt.

Since I was busy at the time with Skylake X, I cannot say how long he ran the stability tests though.
 
CarstenS said:
Raff (he is my colleague) has posted in the 3DCenter thread, that at first, in fact, the card hard crashed (i think he meant blackscreen) with 1,075 mVolt AND at +50% power limit in the ComputeMark subtest MandelS Complex - when run in benchmark mode. He then tried the stress testing mode, and there it ran fine even down to 1,050 mVolt.

Since I was busy at the time with Skylake X, I cannot say how long he ran the stability tests though.
Click to expand...

Is there a possibility, then, that some change on compute side for Vega (including FP16 ) is doing that the gpu on compute intensive test, result on more TDP, more power hungry too when working than in more "normal stressing" ( intensive or gaming ) who dont sollicitate thoses parts the same way ?. ( meaning that thoses part will surely be disabled on the gaming version anyway. )... I see some report, that Vega gaming could be clocked way higher than the FE edition. (1700-1800mhz ).. ( but i cant confirm, first if it is true and not even if this is stock clock )
 
I am not sure, but if packed math and INT8 operations are executed by the same ALUs, I don't see much chance, that a gaming part, where pure compute is of lesser priority, would be able to clock much higher even if FP16/INT8 were disabled for those SKUs.
 
Genotypical said:
its from the forum the author posts in. "Raff" is the one linked to the pcgameshardware.de site

https://www.forum-3dcenter.org/vbulletin/showthread.php?p=11429132#post11429132

there are further discussions beyond that post.
Click to expand...
Ah ok, I've seen it now. So maybe 1.075V is indeed just not quite enough to call it stable at that clock (even though it seemed to run at even 1.05V for compute afterwards, which wasn't stable with gaming). But even at 1.1V, the savings are quite substantial (as expected, of course).
 
mczak said:
Ah ok, I've seen it now. So maybe 1.075V is indeed just not quite enough to call it stable at that clock (even though it seemed to run at even 1.05V for compute afterwards, which wasn't stable with gaming). But even at 1.1V, the savings are quite substantial (as expected, of course).
Click to expand...

Later post says it was stable at 1050mV +50% TDP using a 300 second "MandelS burn test demo" (is this Mandel Scalar from Compute Mark?).
The 1075mV crash may have been just a bug.

He also says that at 1050mV + 50% TDP (constant 1600MHz) the whole system is pulling 35W more than the card with default values.
 
You must log in or register to reply here.
Back
Top