Our STRIX Vega 64 includes reference Vega 64 results at 1440p and Vega 56 was done at 1440p too, I don't see the problem. Only thing missing is the OC of reference @1440p.
AMD Vega Hardware Reviews
- Thread starter ArkeoTP
- Start date
Both Vega56 tests are done by undervolting. They are just different runs, and that's where the small difference comes. Undervolting may sound strange, but it's how you "overclock" vega cards now. Default clocks are set so high already that you can't hit 'em (especially p7 state), unless you undervolt (not even with 50% powerlimit).
If you're thinking of Guru3D's wacky review, I strongly suspect that card was borked in some way, like disassembled and then reassembled again by a previous hardware site, perhaps multiple times. It's like, totally impossible that an oversize monster cooler like the strix would perform worse than the piddly reference blower heatsink.
So, bogus, outlier datapoint. Nothing to pay attention to.
Fake news??!
Unless you are doing benchmark competition.
It is not worth undevolting, Buildzoid was sort of baffled why people wanted him to do this and still could not see the point after his results; made me laugh but then he is more of an extreme OCer/benchmark scorer with an extreme POV.
It is not a problem/issue just it did not help that the complete data like you say is missing for the OC of reference at those resolution, which is not in the original reference review but useful to a couple of discussion points raised in this thread.
Just for reference, can you remember how similar or behind/ahead the STRIX model was to the reference when both manually OC'd at 1440p?
Thanks
Yes, but he uses registry tweak to get powerlimit up to 150%. But as stock card comes as 50% max powerlimit. Highest clocks you can achieve with Vega56 is by powerlimit 50% and undervolting.
No since we never did 1440p OC on the reference
I guess I should had caveated it by saying not exclusive to Vega56.
Did he do that for the undervolting video because I cannot find any information while watching that he did. however it is Vega64.
Usually he does a hard mod for Vega, which killed his FE possibly due to parasitic inductance so the undervolting video was straight after with a new Vega 64.
Basically he measured on the card itself before changes and says it does not reduce the power demand between undervolting and normal OCing due to the relationship with the envelope variables - both will be hitting similar powerlimit if talking about reasonable voltages rather than the extremes.
Regarding Vega56, with powerlimit 50% it is still able to hit 310W average and short burst peak of 376W - this pretty much aligns with hardware.fr
The 310W in-game gave Tom's 1480Mhz, hardware.fr managed in-game 1500Mhz for their Vega56, and they could still OC RAM.
The voltage change is not necessarily due to the max powerlimit but could be how it is using various variables to manage the dynamic boost.
Last edited:
I can't find the video where he has many different power registries , but this video has a lot of good info.
Worth thing noting is that he mentions that you should always max HBM clock, as it pretty much doesn't increase power consumpion at all. 945MHz (64 Default) it's drawing under 20W, and at 1100MHz , it's still pulling under 20W. I was wondering why I get same clockspeed in games, no matter what HBM clock I use.
Tom's 1480MHz , and hardware.fr 1500MHz are quite what you can expect with default voltage +50% power limit. But undervolt that and you can hit 1550MHz
Here are some quick clockspeed data from Firestrike Extreme with Vega 64 LC
Graphics test 1 on loop
Default ~1550MHz
Default +50% ~1695MHz
Undervolted +50% ~1725MHz
Default's are P6 1668MHz@1150mV, P7 1750MHz@1200mv
Undervolted ¨P6 1668MHz@1120mV, P7 1770MHz@1170mV
Vega56 has way more room in undervolting as 64LC has. It has same voltages for P6/7, but with way lower clocks, 1538MHz@1150mV, and 1590@1200mV.
No, we didn't, just the basic, power draw to max (133%) and temp target up to 90 to avoid throttling. That coupled with +175 MHz on core was the max Sampsa got from it, in real testing that resulted 1974 MHz.
Yeah but I am talking about his undervolting video where he was testing to see its pros/cons at the request of his subscribers.
Anyway just checked GamersNexus and yeah they also show it is not power limit but thermal.
What I forgot though was 1600MHz defaulting to 1.2V, but if you overcome the thermal issues it will hit that with 50%+OC, albeit hotter and higher power demand.
Challenge is finding a low enough voltage that is stable for a diverse range of workloads and games to make a worthwhile difference, this is why Buildzoid was a bit baffled on why someone would do this because you are not really power limited however he probably does not care about thermal challenge and power demands due to his extreme OCing POV.
Just to say I do agree with you and others it can be very useful but IMO only for certain situations.
Last edited:
Do any sites/publications even try to undervolt Geforce?
Last edited:
el etro
Newcomer
Bullzoid frequents this forum, with the alias "XeX".
---------------
I'm not a EE or even a eletronics technican specialist(my role is at IT technican, but i work as a eletronics seller) but AFAIK undervolt may affect the lifespan of the card, maybe even as much as overvolt does. The passed from test ASIC(GPU) is passed through a rough voltage testing for to determine what vCore of the GPU will be set for every P-state(in the case dinamic voltage scaling paired witg dinamic clock scaling), according to the ASIC quality of the card. Supplying the GPU a lower vCore will be to put the card in the force situation that a higher vCore can do.
--------------------------
And about custom Vega cards: clockspeed/consumption curve is already too agressive for Vegas, maybe doesn't worth investing too much into better power delivery and cooling for a 300W GPU that is supposed to challenge a $500 GPU on the market.
---------------
I'm not a EE or even a eletronics technican specialist(my role is at IT technican, but i work as a eletronics seller) but AFAIK undervolt may affect the lifespan of the card, maybe even as much as overvolt does. The passed from test ASIC(GPU) is passed through a rough voltage testing for to determine what vCore of the GPU will be set for every P-state(in the case dinamic voltage scaling paired witg dinamic clock scaling), according to the ASIC quality of the card. Supplying the GPU a lower vCore will be to put the card in the force situation that a higher vCore can do.
--------------------------
And about custom Vega cards: clockspeed/consumption curve is already too agressive for Vegas, maybe doesn't worth investing too much into better power delivery and cooling for a 300W GPU that is supposed to challenge a $500 GPU on the market.
The silicon envelope spec is shown down to around 0.8V - 0.9V.
Not sure what voltage the Nvidia P4 is but considering this is a GP104 with same number of cores as the 1080 configured for 50W/75W in its max power setup it is pretty fair to assume it is close to the lower end of the envelope.
The issue is voltage rather than frequency, voltage is what impacts the silicon/GPU/CPU at the extremes but envelope specs do show down to 0.8V-0.9V depending upon node/process so the level of undervolting involved manually should not impact lifespan.
In fact it might help improve lifespan as one is adjusting the upper higher voltage associated with P7/1600MHz to be a bit lower - just saying as this is more likely than undervolting reducing lifespan.
The P4 will have low clocks and low voltage relative to GP104 due to the performance envelope relationship between them, you could end up with stability issues having too low voltage with higher clocks but that should be it.
Last edited:
Note that on Pascal cards, at least vith MSI afterburner, overcklocking also implies some form of undervolting.
Afterbuner deals with a voltage / frequency graph. + 50 Mhz is equivalent to increasing the frequency by 50 on _all_ the voltage steps. Hence after +50 is applied, the card will try to achieve say 1600 mhz using the voltage that was previously required for 1550 mhz. (if it fits in the power limit).
Similar threads
