Gadgetory


All Cool Mind-blowing Gadgets You Love in One Place

Revisit: Vega 56 & 64 at Same Clocks (800-1020MHz HBM2)

2017-10-01
we're winding down coverage of Andy's Vega at this point but wanted to explore shaders at least one more time prior to moving on to other tasks in our initial AMD shader comparison between Vega 56 and Vega 64 we saw identical performance between the cards when clock matched at roughly 1580 to 1590 megahertz core and 945 megahertz HBM 2 we're now exploring performance across a range of frequency settings from 1400 megahertz core to 1660 megahertz core and from 800 to 1050 megahertz HBM - this video is brought to you by the be quiet dark Bass Pro 900 white edition the DBP 900 marks a return to full tower cases equipped with ample harddrive support effective noise damping foam high performance fans and the option to be inverted into an alternative layout learn more at the link in the description below the main reason to revisit these tests is to see if the shaders scale differently at the low end of the frequency range it would make sense that at some point you would see shader scaling that we expect that in some compute tasks as mentioned in the original content piece maybe some kind of production workload you would be more likely to see scaling we would think but we're not testing that today we're testing gaming and so far there's not been scaling and gaming so there's potential that we were either encountering some kind of pipeline bottleneck that we just we don't have a good way to define right now or maybe on the upper end and HBM - if bottleneck where as we all know memory clock matters a whole lot more than core clock so at 9:45 megahertz core maybe we were limited somewhere and today's test is looking at more of that so a couple things here we're not going through 10 50 megahertz on all the tests it's really just the Vega 64 card right now and then the 1,400 megahertz and 800 megahertz low-end will represent the bottom line of our scale testing there's a little bit more territory to explore none of our bag of 64 cards can clock as high as our Vega 56 cards I've discussed this with a couple folks in the industry some overclockers some folks you are familiar with on the media side and everyone I've spoken with seems to have a similar experience to ours which is that so far mods or not we're able to get Vega 56 cards two o'clock higher so this may be something to do with having fewer shaders fewer cores means that you can push the clock a bit higher and while that's exciting from a Vega 56 standpoint I mean it further really drives emphasis that Vega 56 is a great card to play around with not as great as it could be with unlocks bios but pretty damn good and if you're willing to play around with it it's a good purchase especially versus the existing ten 70s but that's not the point today the point is because Vega 64 isn't clocking as high on core as Vega - these six we're kind of limited on how high we can go right now so we can hit about 1700 megahertz for Vega 64 when it's due in 3d mark I go a little beyond that with liquid and power play tables but once you start gaming because of how the frequency behaves with games often that drops to 16 safety so that's where we land with 64 even though our 56 card can do 17 42 and fire strike in 1732 and games stable so that's that's where we stand but anyway let's get through these will start with 3d mark a lot of the emphasis was put on 3d mark this time because we're trying to very rapidly create a line plot of scaling across different frequency ranges it's a whole lot of testing each of these tests was executed a minimum of six times per rep so to speak or iteration and then I mean based on what we saw at that point either more tests were run or we averaged what we got but we're gonna start with 3d mark lots to go through there we'll go through a couple games but that's a lot less than focus today we're opening with fire strike 3d mark fire strikes shows home of the most visible scaling with this new round of tests so keep that in mind going forward will lose resolution of difference on the shader impact as we explore more games 3 mark fire strikes scoring that shows slightly more noticeable shader impact towards the low end when under clock into 1395 megahertz core and 800 megahertz HBM - as we move it to 1590 megahertz and 800 megahertz HBM to maintaining a somewhat equal distance as at 1395 megahertz towards the higher end of the frequency scale the lines merge closer to one another and start dithering around within margin of error territory the mark aided by the vertical error barked we haven't yet pushed the Vega 56 to 10 50 megahertz HBM - but we've included the Vega 64 data point for reference our performance flatlined for Vega 64 once we hit 16 57 megahertz core and 945 mega Hertz HBM - we still gained about 5.5 percent performance between the 945 megahertz and mine 80 mega Hertz numbers on Vega 64 but nothing close to the gains seen earlier in the line plot diminishing returns our encountered ones we hit 10 50 megahertz on Vega 64 so the point where we do actually start losing some performance within margin of error though so it's basically the same relating these scores into a more diluted FPS value we can break things into fps 1 & 2 GT 1 and firestrike relies heavily on loading the GPU with Polly's on primitives and tessellation but doesn't apply much of a compute workload GT 2 increases compute tasked workloads and stresses memory harder and knowing these two facts about 3d marks testing we can better understand why each number behaves the way it does starting with GT 1 FPS the gap is largest at 13 95 megahertz core and 800 megahertz HBM - we're at 97 FPS for Vega 56 and 102 FPS for Vega 64 plus or minus some error in our clocks and again note that these are averaged from several test passes but still there's going to be error in there that gives us about a 5% advantage for Vega 64 the shaders so we can finally see a change here like the first round of tests this is outside of our tolerance for clock error and 3dmark variance it appears that at these lower clocks were seeing a lot more noticeable impact from the shader count increase on Vega 64 than previously but it's still only 5% that 5% gain largely persists to the 1589 megahertz core and 800 megahertz HBM - class where we still see about a 5 to 6 percent gain from the shader count the difference begins to fall within error margins toward the higher end of the clock speeds though overall maintains a slight 1 to 4 percent lead over Vega 56 there are times when the scores were effectively identical as seen in our initial round of tests and in a couple of these other data points later on and that falls within test variance and clock error margins gt2 the second FPS score from 3dmark is more compute intensive than GT 1 these scores are nearly identical across the board Vega 64 does not hold a significant lead in any of these tests except for the first two where we clocked at 1395 megahertz core and 1589 megahertz core both with 800 megahertz HBM 2 and again the first number 1395 and 800 is below what you'd get on Vega for these six out of box so it's not really realistic the second number 15 89 and 800 is also well below Vega 56 is abilities with HB m2 in general and certainly below Vega 64 so given these numbers we kind of know what to expect with games in theory at most you would expect about a 5 to 6 percent difference with those lower clocks but as great as synthetic applications are for rapid fire testing and looking for theoretical scaling and theoretical performance and really pinpointing aspects of hardware they aren't games so let's go ahead and see how that worked out now quick note here again some games will have the higher overclock tests some won't it just depended on which ones were stable on Vega 64 with the higher clocks not all of them could do it and really if we started stepping down from the 16 60 plus territory you're getting so close to our original test value for the core that it just seems kind of irrelevant applying this to games is where we start losing some of that test resolution in terms of looking for differences for honor at 4k plots us at around 47 to 48 FPS average with our original test battery where we were around 1580 megahertz and 940 5 megahertz HBM 2 at 13 90 megahertz core and 800 megahertz AGM 2 for each device performance hovered around 42 average for Vega 56 and around 43 FPS average for Vega 64 note that Vega 64 was also running about 8 megahertz faster it's just this was the closes we can control it given the new booster point no equivalent clock behavior in Vega so we're with an error here there's effectively no difference in for honor at 4k with these clocks looking at just the shaders but then there are a lot of other elements of the card engaged when gaming so it's tough to tell what other bottlenecks might be encountered especially when we're looking at different resolutions like 4k 1080p doesn't change much of this our original number is worth 137 FPS average already and that's nearly dead 137 FPS average for each device the one percent and point one percent values were also close to each other at 1392 1398 megahertz core and 800 megahertz HBM to both Vega 56 and 64 are within 1 FPS of each other we're within error margins here again before honor is exceptionally GPU bound as well so we're not in a scenario where we're CPU bottleneck game ashes of the singularity at 4k seems to be bottlenecking though potentially on the cpu and given how this title behaves that make some sense either that whore was showing exceptionally limited differences at 13 90 megahertz core in 800 megahertz a GBM to the difference between the two GPUs is within margin of error here that said they are also not too distant from the 58 FPS average of the 15 90 microts core nine forty five megahertz hpm to test so these differences are also within margins just barely we can't use this test or much so let's just move on to hell bleed hell-blade at 4k shows no scaling between the vega 56 and vega 64 cards went at 1580 to 1590 megahertz core and 9:45 megahertz HB m2 as discussed previously it also shows no differences at 1390 core and 800 megahertz HP m2 so once again no real difference Ghost Recon wildlands at 4k again shows effectively no scaling at 1400 megahertz core and 800 megahertz HB m2 with our scores sitting within it for actions of a frame of each other at roughly fifteen ninety and nine forty five megahertz we also see no scaling this trend continues up to 16 60 megahertz four and nine 80 megahertz age BM to overclocks up to this point all the testing was conducted a few weeks ago it was before we left for the Linus media group shoot and before the 79 80 XE launch we were about to publish the data of more than ten days ago at this point but ended up holding publication until getting back so that we could add a couple more synthetic tests the other part of our addendum went back to an initial plan of not flashing Vega 56 with Vega 64 BIOS and actually doing it so we went back on that we ended up flashing V 64 onto V 56 and then we ended up overclocking memory 210 20 megahertz on each device just for another point of comparison at the high end using heaven and superposition heaven and superposition were also the new options for the rest of the testing and we added them because we thought they'd be more likely to draw out the differences we performed heaven testing using the extreme preset which is 1600 by 900 it's from 2009 and then we also ran a custom testing at 1080p with ultra preset configurations with anti-aliasing at 8x and dx11 for the API we also manually adjusted tessellation across all options we thought this might give some visibility as to a potential bottleneck from the geometry pipeline starting with the extreme preset which is again 1600 by 900 we'll use an FPS to first show differences and then move on the scores the Vega 56 card at 13 90 megahertz and 800 megahertz HBM to averages at 96 point for FPS after multiple passes with Vega 64 at similar speeds averaging 96.1 these are functionally the same particularly considering we're about 5 megahertz lower on average with the Vega 64 card with the BIOS flash on Vega 56 set to 1660 and 1020 megahertz we scored one 12.05 FPS average compared to 114 point 75 FPS average on Vega 64 that is now a 2.4 percent improvement with Vega 64 which is close to our earlier defined error margins but also close enough to the limits that we can say a pattern is emerging we'll keep this 2.4 percent advantage in mind for now as it may come into play as we collect more data with 1080p testing and test and scaly and we're seeing these results at 1660 megahertz and 1020 megahertz for HB m to the vega 64 card operates with a score of 30 50 point 7 vs 30 17 on Vega 56 or an increase of 1.1 percent the gap widens as we increase tessellation to moderate and that results in a score difference of 2801 0.5 on Vega 64 versus 27 30.5 on Vega 56 or a 2.6 percent improvement on Vega 64 normal tessellation also posts a 2.6 percent difference and that's at twenty six twenty eight point five versus twenty five sixty one point five extreme tessellation has us at two point nine percent improves if they because it could be for showing one of the biggest gains we've seen thus far given the consistency of these results we can safely say that the Vega 64 cards extra sea use do help in this particular game or benchmark rather and we're somewhere between one and three percent for gains depending on the test settings as for the 13 90 and 800 megahertz clocks the scoring is roughly the same across the board we're within tolerance for error and the 5 megahertz clock that prints here unlike fire strike the results appear to be mostly the same at the low-end clocks when testing with heaven and one more final note although we like to do this testing to try and determine differences in shaders and that's kind of the verbage I've been using throughout this video talking about shader differences what we're really testing is Cu differences each Cu contains more than just shaders like texture units for example so other elements of the sea you can come into play before the shaders do but at our level we have no way to isolate beyond sea use so we can't control for individual shaders or texture units or things like that as for superposition this is made by the same company that made heaven it's just been updated for the modern era as such it demonstrates largely the same performance it's built on the same engine as heaven was it's still tessellation and poly heavy so no surprise there we're seeing about a three percent improvement with Vega 64 over Vega 56 with both sets of numbers for this synthetic benchmark and these benchmark they're good tools for demonstrating these differences but if you can't realize them in games it's kind of hard to say when they appear it probably comes down to certain compute workloads we would wager you might see these differences the 3% difference is here more in something like maybe Sniper Elite or that new Forza game or something that uses a lot of async if those types of games tap into these shaders that's kind of what we would wager at this point but either way with the synthetic workloads that really exaggerate things superposition is still showing about a 3% difference so not a huge one for sure and really not that distant from our 2 to 2.5 percent error tolerance given that we can't perfectly control the clocks from run to run so we're seeing some scaling in 3dmark firestrike gt1 which is poly and tessellation heavy and we're not seeing much scaling outside of this this test isn't fully complete I mean you'd have to test every game basically to really concretely and confidently state where and when shaders make a difference clearly there are places where it can if we look at the 3d mark results those are highly repeatable and we see some differences at the very low end of the clocks towards 1,400 megahertz so it's unrealistically low that we're seeing the difference emerge but it does emerge that means that one could expect that some type of game out there programmed similarly to this the application would behave similarly there's also room for production applications to show some shader differences we didn't test those today maybe something with computes although it wasn't reflected in fire strike maybe something with compute heavy workloads would have those differences emerge and ultimately this test is kind of difficult to conduct anyway it's it's very time consuming because you have to log the frequency during the test which changes it changes because of how Vega behaves now it changes based on the game if we set let's say we set a number like 14 22 megahertz target that might get you 1392 1398 Hertz if you're lucky if it's going well or it might get you 1422 just depends on the application and so that might be 1390 an application a for honor or something and the application B maybe he'll blade your at thirteen eighty so there's a lot of work of adjusting the clocks to make sure that they're the same on each card because each card behaves differently just like nvidia gpus now and that makes it tedious to test properly that said the biggest takeaway here i think we've kind of learned indirectly as an aside from this is that vega 56 from our testing with our samples which are not representative of every single sample on earth but a representative of at least 1% of them since that's about the amount of cards we have from vega 56 versus the initial launch in our testing so far it looks like vega 56 is capable of achieving at least slightly higher clocks than vega 64 build Zoid is still exploring this personally his initial findings were similar but he's he only just recently got his vega these six in so there's room for him to discover more things i'm sure he'll do videos on it check them out when he does I've spoken with vendors in the industry who work with these cards and they've seen similar spoken with other media they've seen similar so it would appear that there's a kind of expected hit to maximum clock potential with the increased shader count that's not wholly absurd I there's probably some sort of thermal behavior where you've got higher density of active shaders wiggling her house basically what they do and that's limiting the clock Headroom on 64 on our cards there's room for that to be different for other cards that's what we've seen so that also makes it difficult to test the higher frequency results because with 56 we're doing 17 32 to 17 42 with the mods 64 with the same mods or similar anyway we're doing like 1700 to 17 12 at best and once you get into games that falls down around 1660 and couple games that's not even stable so that's that's kind of interesting that's maybe the more interesting takeaway than the shader differences but for now that's what it looks like so this really just just kind of interesting information not necessarily stuff you can act on if you did want to act on it I would say that once again it looks like Vega Vega 56 is the far better buy than Vega 64 its competitive with the 1070 it's pretty modifiable if you're willing to do things like registry edits you can just double click a file basically and if you're willing to maybe get a better cooler on it whether that's DIY or a IB partner whatever it's you can get pretty far with that card it's fun to play with but it does require playing with it for it to be the best it can be that's not just like a normal overclock it's either under bolt it to tune it down and power consumption or you clock it up and compete more directly in frame rate things like that so it's very much a card where to get the most out of it you you have to enjoy being under the hood there are a lot of people in our audience who do so they get these six it's pretty fun to work with for that but still shader difference don't worry about it right now you can subscribe for more patreon.com slash gamers and access to helps out directly as always gamers Nexus dotnet for the website and you can go to gamers Nexus thoughts Squarespace calm to pick up a shirt like this one or one of our cooler designs like the graphed logo I'll see you all next time
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.