Gadgetory


All Cool Mind-blowing Gadgets You Love in One Place

Ask GN 28: HBM on CPUs, GPU Boost 3.0 Curiosities

2016-09-17
hey Ron and welcome back to another episode of ask a GM this is our first post packs episode so I mostly recovered from being sick at PAX but it did happen this year after six years of going they finally got me we're back another episode there so as always leave your questions in the comments below if you have a question for the next episode Hardware related ideally but if it's games we'll take that too and we'll hopefully address in the next episode so a few things housekeeping to get to before the questions first of all this is the rose we'll : in case we saw this at Computex you may remember it has been impossible to get and basically vanished after Computex but it's here now so we have one of the first models of the case and we will be reviewing it this week so do check back for that the case is from I guess just looking at it from a perspective of Rose Welles case history this is one of the most exciting ones that they've made recently and it does actually kind of keep at the front end of some of the industry trends with tempered glass on the both sides actually left side I'm not too sure about but right sides good fronts good kind of like the Iowa power element this would be the DIY version of that in some ways so we'll be looking at that this week just wanted to bring that up separately and that's because a lot of people said they liked it in our Computex coverage and that it never was heard from again so it does exist this is just cuz it's cool we got one of these at PAX so they give USB keys to media with a media kit on it that normally has b-roll gameplay footage photos screenshots this is the one from the Warhammer team for Dawn of War 3 which we already posted that like the channel but it's just a chainsaw sword basically they're chain sword so just kind of a neat thing that we're gonna be putting on the shelf in the background of some of our DIY video shoots we do so that stuff aside cool stuff aside the first question is from exiled storm definitely seen you around before exiled storm says hey Steve would you ever consider testing with lower end CPUs occasionally to see the effect that they have on your results in a more real-world scenario not x99 system suggesting an i5 or FX chips and then says I understand it's a lot of work so yes we have considered it and we have done it generally let me explain a few things here so first of all the reason we use high-end CPUs for GPU tests is because the basis of test methodology when you're looking at one component is to eliminate potential variables or limiters or throttles or bowel X whatever word you want to use with components so one thing variables you if you're testing GPUs every other component needs to be the same 100% of the time doesn't matter if it's low under high-end just needs to be the same now the part where it matters if it's low under high end is when you start running into issues where the GPU might be bottlenecked by another component in the system so when we say this the idea is that the GPU say you have a 1080 pair with a g3 2 5 8 obviously the 1080 is not going to perform that well when you're looking at the FPS numbers and that's not because the 1080 can't it's because the CPU can't keep up CPU and the GPU both do work for every frame process so what happens in the charts and this may be kind of old news to some of you but what happens is if we did that scenario everything from a 1080 to maybe I don't even know maybe a 970 g3 to 5/8 pretty weak but we haven't done that specific test definitely 1080 and a 1070 would look the same and a GP benchmark if you use a low-end CPU like that now that's a very extreme example so the question here why not an i-5 same thing you just were trying to make sure that in as many cases as possible black ops 3 is a really good example where we sometimes hit in excess of 200 FPS that wouldn't show up with a lower-end CPU now there's a valid point here to be made and it's that the real-world performance of a complete system is still important to look at so the here's here's the decision point we run into as media there's a time budget there's a money budget of course time is the biggest one and with the benchmarks we'll normally do a GPU benchmark and then a CPU benchmark with a high on GPU so they're opposites and you kind of look at the two and extrapolate where you need to be as a buyer the other option and occasionally we do this is a complete system benchmark with a very specific set up normally do that for a size but doing something like benching every card on an FX or an i5 I was actually I was just talking to Luke about this from from - tech tips so the problem with benching every card on two platforms for example maybe a m4 in the future and x99 is we're literally doubling the amount of work and it's just not feasible but it is possible to do something like a standalone CPU test per game Sandlin GPU test per game and maybe a one-off system build quarterly or something like that's that that may be something that I look into because obviously the limit is time it's this is something I would love to do just to give a better understanding of where does a certain CPU and a GPU fall on a chart when paired it would be good for that even though we can extrapolate a lot of that from the standalone GPU and CPU specific benchmarks still good to look at so something feasible maybe a quarterly I 5 plus whatever build or an i-5 GPU benchmark or whatever not something that would we would be doing for every GPU benchmark definitely but something that I will look into because it is requested frequently it's just a time thing and we need more exposure to the content to be worth that investment so you can help out there by sharing it next question is Chen kegui who says as high bandwidth memory being the fast fastest memory that exists today will we ever see a CPU with HBM in the future where I have this in mind but is there a possibility that we can see HBM RAM in the future or how fast how does fast memory is HBM affect PC performance or what benefit can we see from it so Intel Intel's Knights landing is using the closest thing to HBM we've currently seen on a CPU so that would be MCD Ram that is Intel sort of bird of HBM it's not called HBM it's a different technology it's made by micron which is Intel owned micron also makes the gddr5 X as you may be aware used in the GTX 1080 and the Titan X Pascal cards so they don't what they're doing with memory but MCD Rams different from HP M first of all it operates at 400 gigabytes per second for the bandwidth whereas HP M 1 & 2 are a bit higher than that - especially is approaching one terabyte per second depending on what you're looking at for how many stacks you have and things like that because it is per stack secondly the Intel Knights landing stuff for MCD Rams looking at about a 16 gigabyte capacity and it doesn't exist yet it's been delayed a good amount of times finally coming out but it's not a consumer product so that's not something on consumer side but it is coming in two CPUs and it's not just a sort of on paper theory thing that this will actually exist so that's important now the question where I think this gets interesting and we talked about before actually is HP m in the future so this Chen says HP MRAM I think the more likely thing that will happen is you'll see CPUs even on consumer mainstream sides start building HP m to the chip this makes the chip gigantic or I should say makes makes the package of the cpu gigantic as we've seen with Knights landing because now you're accommodating maybe four or eight or six or whatever multiple of two or three banks for the stacks so it's got to be larger for the socket in that way and it works for server boards but not really consumer but in the future I think what we'll see happen initially on the very low-end stuff at least would be a move away potentially from system ram and a move toward memory on the cpu package and that would be interesting because for low-end consumer stuff obviously all this scales as we look into the future eight gigabytes won't really be that much to us in a few years sixteen gigabytes after that will not seem that much that's just how things go but for consumer oriented tasks you don't need huge amounts of memory and hbm's fast it's not cheap but HB m and M cd-rom are fast they're close to the CPU package that means I'm theory you could even shrink the board so this benefits the really sort of use case specific deployments of PCs where it might be used for some sort of pre-built box by ZOTAC or whatever they make their small pcs something like that would make sense so it doesn't exist yet but the theory of something like that happening I think is something that actually could be real not just the crazy talk and the move to HP M or MCD Ram or whatever on the cpu package will start to sort of pull the necessary pull the need away from having dedicated system RAM which is much slower it's physically farther from the chip and that physical distance adds latency that adds access time to every every single transaction it's painting back and forth the memory that takes a lot of time you're painting storage and things like that to depend on what's going on in the system at the time definitely the future it's just a matter of can it be done affordably when can it be done and is it going to be in a capacity that whenever this rolls out is something that we as consumers see as acceptable for our high high capacity applications you're always going to see at least for now some form of DDR ddr2 RAM media productions a good example where we want 32 or 64 gigabytes or whatever we can get our hands on that's not gonna happen in HP I'm affordably anytime soon but the rest of it I think we'll start seeing this rollout with Knights landing so if you're interested in more research knights landing that is an Intel product that is the first one I know of CPU side that will have some form of very high speed memory whatever they want to call it next question is from Henry Silva who says amazing channel thank you I was watching your Corsair MSI Seahawk X benchmarks as well as the EVGA hybrid benchmark videos I noticed the Corsair and MSI model the GPU craps out for one second after being endurance tested over the one hour mark so he's saying we have charts of that after the one hour mark the Corsair Seahawk model card basically the clock rate looked like that and then I slammed down to about 135 megahertz and come back up and recover and do it again repeat and we talked about that actually a couple cards now not just that one so he says this does not happen to the EVGA model do you think this may be due to the EVGA model cooling the vram but while the corsair model overheats that's a good question this is a boost 3.0 issue to my understanding I've been talking to Nvidia about this for a few weeks now we've decided to start just benchmarking cards on our own and try and figure out if there's a pattern I have duplicates of cards now so we can start collecting multiple data sets for one card so here's the here's the issue at a top level first of all the question is at the vram cooling I don't know I don't think so but I don't know secondly boost 3.0 works a little differently from boost 2.0 in how the clock rate responds to different variables so the variables it looks at our thermal power and voltage which I guess is part of power but mostly thermal and power so when we see the Titan X Pascal card we liquid-cooled it and the clock rate increased by 200 megahertz in every application we tested that happens because there was a thermal limitation on the Titan X Pascal by using its stock cooler so it was operating within spec so operating that is defined by Nvidia but the spec could be higher if you put a better cooler on it ie a liquid cooler so 200 megahertz gain that megahertz gain is because of boost it's because the GPU is no longer sitting at a fixed clock rate throttles down very low it should be never below 150 megahertz when idle that's basically off no 3d clock there but when you cool it with this better solution suddenly you get more free more sort of megahertz out of the clock the other issue is power so once we got that extra 200 megahertz on the Titan X Pascal then actually what we were hitting was power availability or voltage for stability on the overclock if overclocking so those two factors are the contributors to how a clock responds to boost but as for the drops we're seeing they are dropping too low based on the nvidia specs so on video assign us should never drop to the 135 115 numbers we're seen in some cases like that Corsair Seahawk card that's far too low that is visible in game and it's normally in the point 1 percent lower than 1% lowest minimally because it is about 1 second so you're not gonna see it averaged out over a 2 hour period of course it just gets smoothed over so what does that mean well it means that and videos answer to us was every GPUs different now even if it's even if we have 4 gtx 1080s all founders Edition we benchmark all of them there will still be an FPS range maybe 1 or 2 FPS on either side of an average between the cards and that's because GPU boost 3.0 looks at all these different variables on the silicon itself silicon's made differently some of its made better than others and those small changes now are dictating and clock rate differences so that means if we review several models of the same card it may benchmark differently marginally but differently and that's important the next thing is that Nvidia was telling us that because of this and because GPU boost dictates that the card throttles down they don't want me to use that term throttles because I guess they define that differently than than we do but let's use let let's use the phrase down clocks the GPU down clocks whenever it can because it's trying to save power this is a big deal with Pascal it's a big deal with Polaris but AMD approaches that much differently and they can't be compared with Pascal and videos approaches every time we can save a watt we do it so they will down clock and that is reflected in the charts a little bit you can see some of our charts are a little spiky they're not perfectly flat lines like the RX 480 when we increase the power target but they are spiky within about 100 megahertz range not a big drop couple percent sometimes so it's those huge drops to 135 that really matter and that were researching the answer to your question is I don't think it's just the vram plate because we've seen this behavior on a few cards now at 1060 included the titan acts Pascal didn't drop this hard it drops but not this hard so I need to research it more and video was helpful to a point and there was a limit to the help I was getting that I could I could sort of I guess integrate into a heart content without just testing myself so we're testing it ourselves now to validate everything once I have a big data set I'm gonna go back to Nvidia and say ok I've done what you suggested now that we're clear that this isn't a software issue or whatever can you tell me more because there's there's only so much that I can figure out obviously on my own there's just there's a lot going on with Boost 2.0 so we will have the report back and see what we find based on their further support hopefully next question FD says hello Steve great video as always I was wondering whether or not a window on a case could significantly impact temperatures and if their material along with the number of windows made a difference do fall glass cases have advantages or disadvantages so alright should throw this into you know FD also says I know you don't review cases often but any insight would be appreciated we used to review cases a lot I'm getting back into it just a there's been non-stop video cards for six months so we are getting back into it with this one soon and then Silverstone soon after that so first of all different types of glass and different glasses this is tempered there normally you're using acrylic so the tempered glass here is four millimeters thick that's significantly thicker than acrylic does the the glass or the window impact cooling though the answer is not really so couple factors where it will impact cooling and where it does is if we have a setup like this where the front is completely glass you cannot take air in through that I know everyone knows this so there's no air intake here in this case like in a lot of modern cases they have moved the intake to this mesh sort of on the Sun or grill I should say on the side so as long as this is thick enough going that way away from me as long as it's thick enough that the fans can still pull air in through the sides and then pull it into the system the coin will be fine it's not going to be it's really it's it will be worse than if the front is mesh but it will not be that much worse that you'll really notice it in most use cases depending on the case so that's the first answer to that question the second answer does the amount of so if you have more more windows that are glass is it going to impact the the temperature only insofar as you will have fewer places for mesh and fans so if your option is to have a glass side panel or a steel one your temperatures are going to be mostly the same I'm sure a materials engineer could tell you that technically one material or the other would potentially conduct a bit more of the sort of ambient temperature within the case but or maybe even seal it in a bit better but for our purposes as users I would be very much skeptical if you saw more than a 1 degree Celsius change if they had probably talking tenths of a degree for the most part so the answer is no unless the alternative is glass versus a mesh window so on side panels especially there haven't been a lot of side intakes lately the 1/2 X this was very popular that case is basically gone now but and I think the Antec 900 did it as well when you have a side intake that is the best possible GPU cooling setup that can exist from our testing even all this front intake it's not better than a side and take straight into a GPU if it's positioned correctly the only place that's not true is with the silverstone raven our vo2 but the reason they're good without side intake is because they have 3 180 millimeter fans on the bottom that are just blasting air straight into the video card or straight up into the system I'd say it is an inverted layout but straight up into the system that it's just a lot of air to compete with 380 millimeter predator fans but yeah I think the answer to that question hidden short is now there's no real difference that one contingency aside stand up or shut up s-- as seems like there's a lot of work put into this channel how do you and everyone else find time the game I don't find a lot of time the game lately I've been trying to play more City skylines on the planes for years I have tried to get into DS gaming or mobile gaming it's not worked for me I can't keep my focus on it I would almost always rather work instead but lately I've been playing Age of Empires 2 on the D s and that's kept my attention so I'm playing a bit of that on planes which there seems to be four of those at least a week for plane trips or her week a month although a week lately I guess it's kind of accurate with packs and all that stuff but yeah it's a city skylines Age of Empires and maybe one day I will get into overwatch but other than that that's about it so thank you for watching as always patron like the post roll video if you want to fund our efforts to buy more time and use it to play games otherwise subscribe for more content comments below for questions next week 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.