Gadgetory

hey everyone I am joined by David Cantor today from real-world tech he's a technical analyst we've spoken with David before on I think we last spoke David about CUDA cores and whether or not it's accurate to call them that yes yes with with with my verdict being no right yes David is back today to answer a question a lot of you have had in the audience which is I'll just read it actually this is from a viewer from ask GN who said what do you think of Intel struggle at 10 nanometre and AMD moving to 7 Ana meter with TSMC and then further stating Intel CEO had to resign over the 10 nanometer struggle and they may truly be in trouble with the constant 10 nanometer delays so that's our subject today before that this video is brought to you by Thermal Grizzlies high-end thermal paste and liquid metal Thermal Grizzlies cryo knot is an affordable high quality thermal compound that doesn't face some of the aging limitations of other pastes on the market cryo not has a thermal conductivity of 12.5 watts per meter Kelvin focuses on endurance is easy to spread and isn't electrically conductive making it safe to use on GPU dies thermal grizzly also makes conductor not liquid metal which we've used to drop 20 degrees off some temperatures than our dee-lighted tests by a tube at the link in the description below where do we even want to start with this one that's kind of a big topic I mean I think we can both agree that definitely Intel is delayed on 10 nanometer I think that's pretty known so and I just saw a new report this morning that claims Intel's targeting systems on shelves for holiday season of 2019 for 10 nanometer so with all of this known I mean what can you walk us through I guess even just the top level of what's going on when was this supposed to come out why is it delayed yeah so I think if you look at the history of of sort of the manufacturing side right which is the the lowest level of you know where we start making our architectures intel on the industry had always been on the the Moore's law cadence of moving to a new node every 24 months and so if you look at it in that context you know we expected to see 10 nanometre out in 2016 and you know Intel talked about how oh well you know 10 nanometer is a bigger shrink from 14 nanometer than a normal node it's maybe more like you know 1.25 X of a node or something along those lines and there's truth to that but the reality is that you know either no matter how you slice it you know Intel originally wanted to have 10 nanometer in production in 2016 and it's looking like it's now more like 2019 so that is a huge miss but it's also true that if we go back 14 nanometer was delayed a little bit and that was more like you know rather than being 24 months was more like 30 to 36 months and just the delays got worse at 10 nanometer and things sort of seemed to have snow balled a bit but one of the interesting things is when you look at Intel's 10 nanometer process right it is very much comparable to what the foundries are calling 7 nanometers and you know sort of one of the complications is that when Intel move to FinFETs they both shrank from 32 nanometer to 22 nanometer and added FinFETs and when we talk about process technology there's sort of like two really important dimensions and one is sort of what's the performance that you get out of it you know how fast is your chip run how much power is it burning and then there's the density and Moore's Law actually only really talks about density but if you want to make things smaller power does need to go down otherwise you end up with you know sort of you know a thousand watts in something the size of a penny would be very dangerous okay yeah yeah so uh when Intel went from 32 to 22 they added in FinFETs and it was a really great process but at the same time the foundries moved from 28 nanometer to 20 and they did not add FinFETs and that in many respects had some very significant drawbacks because that was also the first point in time at which the foundries had to use double patterning to draw uh the the finest features and and so for those who aren't intimately familiar with it you know when you want to set out a bunch of metal what wires um there's a limit to what we can do with our current lithography and if you look at what goes on today it's almost like you know these these amazing impressionist paintings where you're drawing out you know really fine details like the eyes or a smile with these huge brushstrokes and that's effectively what we've been doing with double patterning and now you know techniques like triple or quadruple the expensive tool in the fab and so what happened is when the foundries went from from 28 to 20 you know I think you know you might recall from talking to Nvidia and AMD that are they were gonna skip it and it's because all it did is it gave them better density but also drove up cost without giving much in the way of performance and they said well hey you know I can't sell products with this we're gonna wait for FinFETs and so then when FinFETs came along you know all the foundries marketing guys got together at all we think this is such a huge improvement we're just gonna call this a totally new note even though the density barely improved so the reality is the foundry guys have been you know blowing smoke on their node size for a really long time and you know for whatever reason people have let them get away with it and you know even internally when you talk to folks at you know real companies they'll always say oh it's the foundry seven nanometer and Intel ten right and and from a density standpoint those those two nodes are more or less equivalent although it's highly likely that Intel will have better performance so this is I guess to to kind of recap an important point here that I'm hearing it it's you can't just strictly compare ten nanometer versus ten nanometer from one foundry versus another or one one company's design versus another and they're not a not all versions of 10 10 nanometer are created equal that's exactly right and it sounds like density is the the major consideration there as well yeah I mean it's both density in performance and look at the end of the day you know your process technology is incredibly complicated there's so many different knobs to turn and you know trying to boil that down to one number to 10 nanometer or seven nanometer is is a vast oversimplification and you know in the past I think it was actually a pretty useful shorthand but I think over the last you know call it a 5 to 10 years you know some people have frankly been been less honest about what a nanometer is then than others right so when when we're talking about I guess this I'll read a a note that I saw from extreme tech earlier when researching to that I think speaks to your point where they said while tsmc and Samsung have both been shipping 10 nanometers for quite a while Intel's 10 nanometer node targets feature sizes that correspond to what pure-play foundries are targeting for 7 nanometers so yes so I guess the way to put it is that if you were to take so look I mean AMD is doing AMD's n processor is you know look it's a different architecture then right then then Intel's skylake or canon like Rice Lake but I think the right way to put it is that from a density standpoint if you were to take an intel ten nanometer design and reoptimize it for the design rules at a foundry seven nanometer process you would come out with something that is probably similarly sized now the performance might be different and getting into performances is a much more complicated question I think my expectation is that Intel will actually at ten nanometer have better performance than the pure play foundries at seven but at this point you know things have sort of gone so sideways that you know even that is maybe a little bit unclear right right so what when I was reading some quotes from Intel's now former CEO Brian krzanich he was citing ten nanometer delays as at least partially due to a lack of a UV which you mentioned briefly earlier so the first question I have I guess is let's boil down what is extreme ultraviolet lithography and yes start from there right so so today most of the lithography that is done in production uses 193 nanometer light to draw different features you know whether it's your transistors or your wires or other components and you know getting back to that point I made about you know sort of impressionist painting when you talk about like just wires they are currently you know the industry is in production with you know 50 nanometer ish pitches so you know that's like you know a quarter of the wavelength of the light and so the fact that we're even able to make that work is is pretty magical and amazing and I think you know if you've gone back to when people were originally thinking about EUV they'd say oh that'll never work an Eevee by the way is is not new technology it is very old I mean I think you know UV has been the next thing since the late 90s and you know I think it has a 25-year track record of being three to five years away from production but but the core observation was that like hey look you know at a certain point in time we're gonna get two feature sizes that are small and are below 193 nanometers and we're not going to be able to do it with standard lithography so let's go work on something that has a smaller wave life and so EUV is 13 and a half nanometer wavelength light so you know that you know if you can't want to think about it going back to my painting analogy it's like going from this you know impressionist painting where you got a big brush and need to you know eke out some fine details - all right now we get you know a real sharp pencil and we can sketch everything in and then fill it in and and we're gonna get some really nice lines in there and so that that promise is very attractive to allowing people to keep on scaling density without running into some of these yield issues that Intel has been running into and and and you know that's the promise the reality is EUV machines introduce all sorts of other problems on their own you know they've been delayed for 20 odd years and you know it's a different technology so it's absolutely gonna have a lot of advantages over standard immersion lithography but it will introduce problems of its own and so you know what Brian was really saying is that you know in the absence of EUV intel has to go with what you know they think can be put into high-volume production which was obviously you know sort of more standard lithography and it was just the physics are getting harder and harder and her every year and Intel made a lot of big aggressive bets with five with with 10 nanometre and you know some of them didn't work out and you know that has led to a situation where you know sort of as we speak apples in production with TSM c7 nanometer process and they're gonna be producing you know they are rumored to but we all know that you know there's gonna be a new iPhone in September or October and that's gonna have TSM c7 nanometer chips and for you know a company like Intel which was you know started by Robert Noyce and Gordon Moore and you know Andy Grove was was not quite a co-founder but you know you know for all intents and purposes was there from the beginning you know that is embarrassing Intel is nothing if not a manufacturing company and and you know Brian was not fired because things were going smoothly and the company was kicking ass I mean this the the stock is at an all-time high so in that sense you know the company is doing well but you know for a company that is a manufacturing company to have lost their manufacturing edge over the course of the last you know five to six years you know that's I yeah i think that's really at the the root cause of what's going on right so do you think that looking at the statement of a lack of EUV being a major reason for the delays do you think there are other major delay causes and lack of EUV yeah so there there are other things that Intel changed in 10 nanometre that are likely a a contributing factor in the delays uh you know I will say that that ultimately within Intel there's probably a very small number of people who actually know exactly what is causing the problems and you know if the CEO of Intel doesn't want to say on an earnings call what the issue is you know other than in broad generalities you know you can be sure that that that the rank-and-file employees definitely don't want to say to any Outsiders right right so um one of the points that Brian brought up was that Intel is doing something called quad patterning where to form certain patterns you will expose you will use self-aligned kuan patterning or even even greater techniques to to do multiple exposures and that just gets really complicated and so that's that that is a problem that would absolutely be solved by EUV almost all of the dimensions where Intel is is drawing things can be handled by an EU v tool the challenge is that those EU v tools are just not ready for high-volume manufacturing and you know both Intel and tsmc have been really clear that they're they're you know they they need to see significant improvements before they can put a UV in production right and just to to briefly explain talking about multi patterning or quad patterning what is the major like you're going to describe this to me in an elevator what's what's the major difference between that and UV the in terms of how how you would use multi patterning to make make a processor yeah so so in the absence of a vuv when we're talking about these really small pitches you know if we're gonna use 193 nanometer light sort of the idea is let's say we want to get some lines that are a hundred nanometers apart just just to make the math really simple then one way to do this is to draw set of lines that are 200 nanometers apart right and then draw another set of lines that is also 200 nanometers apart but they're offset by a hundred nanometers so you have one set of lines that's it like you know 0 200 400 600 and then another set of lines that's it 100 300 500 and so forth so by introducing those you can sort of get the effect of much denser patterns but you have to run it through your lithography machine twice into all of those related steps twice and uh you know the reality is that this is real life right nothing's ever perfectly aligned so maybe they wander a little bit so maybe one line is really it's not at zero it's it like +5 nanometers and you know maybe the the other line is not at a hundred its it - you know 5 so 95 so those lines might be a little bit closer than you thought and so some of those things ultimately you know eventually you do this enough and statistically something may become challenging and not behave in the way that you want especially when you're doing you you know some technique like that 4 times so you know it's basically a trick to run a given image through the the immersion lithography tool twice to get a finer pitch that's below the wavelength of the tool and then that seems like it would impact the yields based on what you're saying if things don't go through perfectly each time I would think that you have to throw away whatever processor went through there um yeah it has it has a bit of an impact on yield I think you know for something like the example I gave you a double patterning it's really actually much more about um cost okay all right which is that it takes more machine time so they can make fewer processors on those machines I guess yeah exactly in that amount of time yeah and and the other thing is I should also say that this isn't unique to just Intel TSMC Samsung and Global Foundries right the DRM guys also have to deal with this right and so you know I was actually at a conference pretty recently where micron said yeah you know there's a UV out there but we don't see that as being cost-effective for DRAM you know we're gonna need to do quad patterning sooner or later and that's going to be a much better bet for us and so you know this is really an industry-wide problem is as we scale down these really extreme features how do we how do we do the imaging but to loop back to your higher-level question you know there is a lot of new stuff that Intel put in there ten nanometer process and the bottom line is that you know it looks like several of those bets didn't work out it does look like you know if you V had been ready in time that it would have certainly made several things easier but I'm not sure that if EUV had been ready in 2016 that everything else would have been working a hundred percent right so then I think my last major question for this one we can we can always push more of this discussion at the future videos because there's a lot here clearly to my last major one for this is sort of at this point Intel's 10 nanometer processes it's about as fabled as half-life 3 where you start questioning whether the the payoff if there ever is one and there will be one for Intel unlike maybe valve the question whether the payoff will will be will live up to the expectations of at this point consumers for multiple years now I mean we're looking at 2019 for a realistic launch from what Intel is now saying Andy is real competition at this point it's not it's not just bulldozer anymore so so at this point there's Khan texts to this where it actually feels like there's some sense of urgency to ten nanometer so can it live up to expectations I mean do you think that when Intel finally launches a 10 nanometer product are they still going to be competitive are they going to lose some of their competitive edge but still keep some advantages what are your expectations there yeah so I think that's it that that you know yeah that gets to the heart of it certainly and look you know if AMD is out with seven nanometer right there you know and Intel is still mostly using 14 then it's clear that you know in certain respects you know particularly with respect to density AMD is going to have an advantage and I think the you know really the multi-billion dollar question is okay well what does that translate into at in terms of product performance into into cost and you know how does how does in Intel's 10 you know sort of look when it eventually comes out you know a lot of that is unknowable but and certainly I have some highly educated guesses but you know I think the bottom line is it does appear like this is going to open an opportunity for AMD to execute in you know 2019 2020 timeframe where they may you know gets a market with a new product on a process that's more advanced than Intel's and wait where can that help them right where are they going to go first and so I think it's quite possible that we could see that giving a real boost to some of the AMD products in that time frame but if it's you know if AMD effectively comes out with a seven nanometer part you know that consumers can buy you know three months ahead of Intel's 10 nanometer parts yeah that's I think that becomes more of a PR blow then than anything that really matters and so you know I think there's a lot of uncertainty around those those time tables and then also what it what it how it matters in performance I mean I think Intel you know has been truthful and that they have really ramped up the performance of their 14 nanometer process and you know so TBD on how other folks can match that right I mean I think you know from an overclocking standpoint right in your a bit of the expert here not me right would you rather be overclocking you know with an Intel 14 plus plus chip or you know an AMD right yeah which part of that will come down to well I was just run over overclocked an enthusiast standpoint you're looking at how much Headroom is there to really gain before you write really crazy things right yeah so and that'll just come down historically right now and these really pushing towards sort of the limit just out of the box they're pretty close to the high frequencies at highest frequencies you can get so I find some of the AMD processors presently to require a whole lot more cooling and effort and everything to get bigger more impressive gains than a couple hundred megahertz but that could obviously change significantly with the next next architecture as AMD kind of maybe it settles into some of their performance uplift and maybe doesn't need to push quite as high as a frequency straight out of the box or something like that so well we'll see how that goes but yeah that's that's an interesting question - yeah and and you know I think that's yeah I mean that's really what it comes down to is you know how much performance is AMD going to get out of seven nanometer how much more performance can Intel get out of 14 and then you know sort of what does that translate to at the the product level and you know look I think you know the other question was ok well his Intel's 10 nanometer gonna live up to the long-delayed hype and I think it sort of depends on if they're able to get into production with what they originally defined is 10 nanometer or not you know there have been some rumors that they're looking at making you know maybe a slightly easier to manufacture version and that that's gonna be what's out the door and so they might give up some of the really impressive density gains the they reported um you know in favor of getting something that they can actually put into high volume because you know I mean like right today you know Intel has the token ten nanometer cannon lake part and in many respects you know the guys who I feel worse for are the whole Canon Lake team who you know did so much hard work all right to get a product out and you know and this goes for a lot of the people who were designing products on ten nanometer and you know the you know it really was sort of the manufacturing that that that upset things and so yeah there's a cannon like part that I think in theory is being sold to Lenovo but for all intents and purposes you know Intel's Intel's ten nanometer processes is not there today and so you know will it be there in the future absolutely and I think the question is okay well what do they you know is it going to be the original ten nanometer process is it going to be the ten nanometer + process for ice lake or is it going to be some sort of sort of you know scaled back a less aggressive ten nanometer process and you know my crystal ball doesn't work quite that well yet not yet we'll keep working on it yeah thank you for joining me so David Cantor has I suppose will send people to your website real-world tech comm and David has some some excellent write-ups over there that are on the technical front I think you've done some topic thought some content pieces on ten nanometer dating back to when it was probably originally discussed so yeah you know I think actually if you know if you're really interested in looking at some of the details the gory details of semiconductor manufacturing I think earlier this year I wrote about Intel's twenty-two FFL process which is a 22 nanometer process that actually uses some of the techniques that Intel pioneered in ten nanometer and sort of just back ported to a less expensive process technology um that could be a great thing to read you can also follow me on Twitter I'm the canter and yeah i mean i i'm you know we will hopefully have more write-ups at real-world tech I know my articles are not nearly as frequent as yours yes to be fair they go much greater depth takes a bit more time I think yes check out check out real world tak calm it's got some really good stuff on there another brief example is when David went into some of the Maxwell or Pascal tile back as well yeah yeah yeah tile based rendering so all good stuff thank you for joining me again and everybody if you want to subscribe to us hit the link below or you can check out David's website will link that below as well go to stored I came in sexist net or patreon.com slash gamers Nexus helps out directly and David hopefully I'll talk to you again soon for another topic on this absolutely who's a pleasure thank you alright we'll see you all next time you