Gadgetory


All Cool Mind-blowing Gadgets You Love in One Place

Undervolting Vega FE: Fixing Performance & Power

2017-07-17
under volton vega was one of the top requests on our user-submitted to-do list for this card and it yielded some of the most interesting results and data set this feature benchmark shows how we can significantly reduce power consumption on Vega while yielding better performance from the FE card not even the same just straight better performance better power consumption and better thermals before getting to that this is brought to you by our patreon backers at patreon.com slash gamers Nexus who helped us in funding these at user requested benchmarks you can go to patreon.com/scishow and access tell us that directly or if you prefer you can grab our brand new shirt design which has a link in the description below so this benchmark was requested a lot for Vega because in past and the card generations like even the rx 500 Polaris refresh it's been possible to undervolt the cards or reduce the voltage going to the core and achieve either the same or slightly better performance in terms of clock stability while also potentially reducing power consumption as opposed to for example a straight plus 50% off set without under voltage in which scenario end up destroying a lot more power stabilizing the frequency but it's at the expense of extra power and extra heat because extra power so that's why this was interesting to people and it was interesting to us as well hence running the test but a few things here just to get everyone on the same playing field power target and voltage are sort of tied at the hip but they're really not the same thing so you can increase power target and decrease voltage and they're not really going to conflict with each other what's happening is these two different items power target limits how much power is permitted to go to the GPU so that might be something like 20 to 21 amps and the stock state for example and then of course you have thermal targets as well like 80 or 85 C which will limit you thermally so you have two different variables there where you can be constrained thermally or constrained and power consumption or power availability to the card so if you increase the fan rpm on this to something crazy like 100% you eliminate that thermal restraint but you now are facing a power constraint so that's that's part one part two voltage is what's applied to the core in this case technically you can change the memory voltage as well we're just looking at core voltage today more voltage means more power consumption so that is problematic potentially if one you just don't want to draw as much power with really no need and two it means more heat so in the stock state we're actually allowing less power to go to the core because you're at a zero offset so you're 100% power target or hundred percent power consumption as opposed to 150 for example so you're allowing less power to go to the core and you're running in this instance a higher voltage than is necessary and the cards have DPM States so they go I think it starts at zero but they go up through seven and the seventh state is basically the highest boost possible so out of box for this that's 1600 megahertz and out of box the voltage is 1200 millivolts so 1.2 volts the next state down is 1528 megahertz and the voltage is a bit changed as well and then below that you start getting into the territory of 1440 megahertz so in the initial reviews when you saw people saying that the card wasn't sustaining 1600 megahertz it was probably 1440 to 1528 because those are the two options in the pre-configured DPM States in a.m. these drivers and BIOS for the vega card and the result is that you end up with lower performance because the clock is throttling down to those two states and again it's only one or two reasons for this it's either the power or the thermals so under volting is a way to solve some of that and that's what we're getting into today so targeting methods and problems first in case you want to do this you should be aware of these things first of all problems what man is not perfect it has a lot of bugs some people have not run into the same ones that we are but that's because they're doing things differently if for example you go in for the overclocking and voltage tools and you set them all to the same number so sixteen hundred one through seven and twelve hundred one through seven then we have seen especially with other changes in there like fan speeds issues where HBM drops to 500 mega and that will really hurt your performance so be aware of that basically I don't know the exact answer as to when HBM 2 drops to 500 megahertz it just seems to kind of happen as you're tweaking stuff in there and I'm not sure which change causes it but just be aware that it can happen so if you go in with a plan to undervolt like this just keep an eye on HBM speeds after you're done to make sure you haven't actually worsened performance though your power consumption would probably get better so that's one thing to be aware of what tool recommended to me by build Zoid is pretty good but has a lot of problems as well it has the same 500 megahertz memory bug and it seems to be caused by the same things but what tool is pretty easy to work with if you prefer that and then finally the fan targets are wrong so you can go in and set a fan target in watt man or watt tool of say 3400 rpm and it'll add a couple hundred rpm on top of that so it'll go up to like 3700 for example close enough though close enough to use but just something to be aware of and then finally for our testing we have three configurations to look at primarily one is stock Vega so 0 changes at all after a driver install the second one is stock Vega with a plus 50% power target and a fixed fan speed 3400 rpm in Hua Man which is like 3700 in reality so the important part plus 50% on the power and then the last one is stock Vega in this configuration plus 50% power under vaulted to whatever was the most stable and with the fan speed also at 3400 so stock with plus 50% and stock with plus 50 percent under volted are going to be directly comparable thermally whereas the stock one will not be because it regulates its own RPM and so will be different so that's the setting we're using this will allow us to ensure that we're not hitting the power lot wall ensure we're not hitting a thermal wall and leave us with an understanding of how much voltage is required to remain stable for our voltages we tried a lot of them 1080 just wasn't going to cut it 10 80 millivolts not good enough we then tried 1090 millivolts down from 1200 mind you and 1090 was stable in a lot of the titles it was stable and firestrike basically ad infinitum it was not stable in for honor so for honor we had to actually go up to as high as 11 20 millivolts and then some other games required 1100 or 1090 so it's not perfect basically again if you're planning to do this be aware of that you will either have to take the highest stable voltage like 1120 for example in our case for all titles or you'll have to get used to the idea of changing voltage for each application or setting profiles I suppose you can do as well so be aware of all that but we can get into the charts now start with current draw at the PCIe cables only not counting the PCIe slots the completely stock card starts off drawing about 268 watts but as we approach the 400 second mark the card starts spiking hard between 17.7 amps and twenty three amps this behavior correlates with clock throttling which we'll show in a moment and is precisely why we've been saying that Vega Fe air can't hold its advertised 1600 megahertz boost clock out of the box its power limit and cooler are simply insufficient the cooler can do it if exiting the fan profile and going to high DPS with your own fan profile but this is where it sits out of box that gives us our baseline the next move is to get the frequency to hit 1600 megahertz constantly so we increase the power target by 50% and set a fixed fan speed and solve for the thermal limit absolving Vega of both of its main limitations anyway at once the red line is the result a new problem emerges the normals in frequency are now under control but the PCIe cable from the power supply is putting out 30 amps at the time averaging about 28 to 29 amps overall that's about 344 to 370 Watts down the PCIe cables and it's going to start generating more heat in the card as a result but it's also just kind of wasteful power consumption finally our under voltage line emerges the blue line represents an under volt of negative 110 millivolts so we've dropped from 1200 to 1090 millivolts current is now 23 amps for power consumption 283 watch at the PCIe tables that's about 15 watts more than the stock set up that struggled to maintain 1600 megahertz about 87 watts lower than the power offset setup that sustained 1600 I guards successfully and should lower thermals as well let's go to that chart our orange line again represents the stock auto configuration which runs a fan curve that isn't aggressive enough a voltage that's too high and a power budget that's too low the worst of everything the result is constantly hitting the thermal limit and potentially throught well definitely throttling observed at the 85 C mark though there are a few spikes to 90 C it tends to stick at 85 applying a 50% power offset and fixing the frequency to 1600 megahertz properly temperatures are about 73 Celsius now but the fan is 70% to control for the thermal variable for under bolting so noise levels are somewhat unbearable at 60 DB a versus the auto noise level of roughly 50 DBA but there's room to drop the fan speed with a lower voltage because less heat is generated as less power is consumed we just stuck with these numbers because we really wanted to make sure there was no thermal limitations that would mess up our undervolt in tests the more appropriate comparison would be our blue line versus our red line has these two are tested with the same settings aside from just one variable voltage to the core the fan speeds the same the 50% power offset is the same and the voltage is the only difference at a 110 millivolt reduction on the under voltage card where we perform at 53 to 66 Celsius for about a 7 to 10 Celsius reduction from the card operating at 1200 millivolts that's pretty good so far the last question is a frequency plot in frequency the orange line that shows the stock out of box configuration for the Vega frontier - an air-cooled card we're throttling hard and only rarely achieving 1600 megahertz the regularity of which 1600 megahertz is achieved diminishes significantly as time goes on largely due to thermal constraints in a default fan curve we tend to be operating at DPM power state 5 to 6 rather than state 7 which would give us full performance the red and blue lines converge on this chart as increasing the power target and removing the thermal limit gives us a perfectly flat 1600 megahertz curve closer to what advertised on the box that said the red line is polling 344 to 370 watts to the PCIe cable so it's a little aggressive and may not be worth the power and thermal load overstock under volting however permits 1600 megahertz to remain possible 50 percent power offset possible and thus permitting 1600 megahertz and still draws 87 watts less powers on the redline with only 15 watts more than the orange line that's pretty good trade so that proves the theory and shows that under bolting is working at least once you learn how to deal with the applications and their bugs and it looks good so far now obviously a maintain 1600 megahertz or sustained versus 1442 1528 fluctuating is going to perform better but just to really drive at home we're going to perform a few benchmarks on 3dmark ghost recon and doom and a couple extras will be in the article linked below but the point isn't to re benchmark the entire test suite we're just going to give you a worst-case a best-case and then a synthetic test so that there's a pretty wide spread of what under Bolton does in the real world now additionally we are leaving out spec view perf again not fluidly the point of this if you want spec you perf performance numbers check our hybrid Vega card video with the results as that has the very top end of spec you perf performance once this thing is overclocked as far as we could get this particular card fire strike ultra starts us out the Vega Fe err card when completely stock ran a graphics core of 4906 with our 50% power offset cards both operating at around 50 to 70 graphics core this includes the under-voltage card which manages about a nine to ten percent performance uplift over the stock card here's the crazy thing again we're not overclocking to achieve this all we're doing is making more power available while reducing the voltage which Nets a marginal power consumption spike at the trade of more consistent and faster frame time that's a pretty good trade for 15 watts and as far better than the 87 watts of the power offset without under bolting card that we did for point of reference our hybrid EFI overclocked performed at 57 74 which is 7% faster than the under voltage card kind of puts into perspective just how far are under bolting and over and we'll get you to begin with fine spy gives us a gain of about seven point six percent from the under-voltage card over the stock card with our hybrid OC gaining another nine point six percent on top of that additionally the drawing significantly more power at around thirty three apps we have a few more fire strike charts in the article link in the description below those interest you has four games some experienced instability at 1090 millivolts and had to be moved up to 1104 honor was particularly unstable and required a core voltage of about 11 20 millivolts so these aren't perfect and can't all sustained at 1090 but we've got to spread and overall you're looking at better performance versus stock anyway let's look at Ghost Recon first at 4k and with very high settings the under-voltage AMD Vega err card performed at 41 FPS average with lows close by at 37 and 36 the stock card with no modifications whatsoever operated at 37.7 FPS average resulting in a performance uplift of 8.8% from the stock card this uplift is because the stock card cannot maintain 1600 megahertz without power offset but again a power offset without over-voltage increases your power consumption by 80 to 90 watts thereby increasing the thermals that the card deals with this under Bolton and overpowering appears to be the best approach to extracting more performance as we've done here with doom using Vulcan async compute which is enabled without anti-aliasing these days and rendered at 4k the Vega under volted card operates at 70 1.6 FPS average with low-end frame times also improved over the stock card our average FPS improvement is about 11.5% in Doom following the trend of Doom being a somewhat best-case scenario for AMD on a routine basis the performance table is tremendous here when considering our minimal power consumption increase and better overall control of the card so that should hopefully answer at least most of the question of under-voltage on Vega Fe thank you for the suggestion for that one a few notes here for the end of this first of all obviously this isn't a perfect solution no matter what 10 90 millivolts is great but it doesn't work everywhere so you've got to add according to whatever the software demands you have to add the power consumption to account for that difference separately at this point it's really kind of like buying a project car if you're doing these kinds of modifications hopefully it's because you enjoy them because otherwise you're dealing with buggy software that'll set you back until you figure out it's weird kinks and anomalies and most of those just save you time seem to be if you set D p.m. states 1 through 7 all to be the same frequency and voltage which is not an issue on the other cards but is on this one because of some weird fury left over code but once you pass that you still have an issue of tweaking the voltage to be a minimum voltage required for stability for each application if that's not appealing then you you're either left with run at stock or run it with the the highest lowest voltage possible to sustain everything so below stock but maybe something like 11 22 11 40 millivolts rather than 1200 that point you're not really saving a whole lot it's certainly better than 1200 when it's not needed but it would it would make sure that pretty much every application works within another note not all GPS are the same ours may undervolt better or worse than others including the ones that you buy so keep that in mind too you probably don't just copy the the numbers we got because it might not work on yours now as for why this is the case it's kind of it's weird this is something that is somewhat consistent with AMD GPUs lately where you can undervolt them and improve performance and at least one aspect and as for why that would be the case my assumption would be that this is something where Andy is picking a voltage that they know will work and keep the card from crashing because that's what happens if your voltage goes too low on all of their silicon so it might just be some kind of safe net of we know 1200 will keep all these things afloat in all applications so we're just going to set it here whereas if you're fine tuning the thing you have a lot more room to play but no company is going to do that so the other option is they didn't have enough time to test it and they pick the highest voltage that they could set without being ridiculous and knew that it would work and keep things stable at which point it more time and testing but not really sure which is the the reason for the higher voltage than necessary but either way you have room if you buy either this which probably no one in our audience should buy or rx Vega which is more likely to be a viable candidate for you it's going to apply there too so you have room with the Vega architecture and cards to potentially undervolt and improve performance now rx Vega is not out yet well test it we'll see if we can under bolt that one as well and if they've set the same conservatively high voltage curve volt frequency curve on that card as this one then there might be room for power consumption reduction on rx maker but all that is out in the future so stay tuned for that subscribe for more as always you can go to patreon.com/scishow exit helps out directly and then we also have our new shirt design which is the GN anniversary edition shirt all the direct link to that in the description below or you can go to gamers next with Squarespace com thank you for watching I'll see you all next time you
We are a participant in the Amazon Services LLC Associates Program, an affiliate advertising program designed to provide a means for us to earn fees by linking to Amazon.com and affiliated sites.