Intel has a plan to go beyond 3nm chips | Upscaled

intel has revealed more details of their plan for unquestioned leadership and there's actually a lot of cool tech here now if they can just hit those deadlines [Music] welcome to upskilled our explainer show where we break down the tech that makes your gadgets work a few months ago intel appointed a new ceo and unveiled a new business plan that included building a bunch more chip making fabs launching a new contract manufacturing business that will make custom chips for other companies and it laid out a road map for a return to on question leadership in the chip making space problem is they didn't really say how that was all supposed to happen but we finally have a few more details of their plan we practically built this show on explaining all the ways that intel has gone off the rails and for more information on that check out well seriously like pretty much any other episode but the briefest summary is that after a rocky few years in the early 2000s intel rose to become well yeah the on question leader in the chip making space starting with the core 2 duo in 2006 and the first i7 processors in 2008 intel was on top their cpus were faster and more efficient their manufacturing process was more advanced than the competitors and their hold on the market was extreme enough to be legally dubious but things started to slip around 2015 intel started to miss deadlines and targets and their pace of improvement slowed way down at the same time tsmc in taiwan began to emerge as a cutting-edge chip maker and a major competitor and amd with their zen cores first released in 2017 eventually reclaimed both the performance and efficiency lead so what is intel's plan back to the top well it's a mix of new technologies coupled with an aggressive road map that will try to bring chip making back to the breakneck moore's law fueled pace of the 90s let's see what's new first off intel is renaming their nodes these are the set of manufacturing processes that chip makers use and moving to a new node generally involves building smaller transistors which can turn on and off faster and use less power this is one of the main ways that pretty much everything we use in the modern world gets faster and more efficient moving to their 10 nanometer node took intel years longer than they'd originally planned though it is finally in bulk production but in the meantime competitors like samsung and tsmc have moved to 7 and even 5 nanometer nodes but remember while these numbers used to measure something specific namely the width of the transistor gate these days they're measuring something in reality the node names are pretty much just marketing and in terms of density the number of transistors you can actually cram into a given area intel's 10 nanometer is in fact ahead of tsmc's 7 nanometer for a long time intel claimed that anyone who cared what these numbers meant new enough to know the difference but it must have annoyed them because they are renaming their upcoming 10 nanometer enhanced superfin design to seven nanometer this node should deliver a 10 to 15 percent improvement in performance per watt purely through changes in the materials and transistor design itself it doesn't actually do anything to improve the density i think it's clear that intel didn't really want to still be talking about 10 nanometers while its competitors were talking about five expect to see these new seven nanometer transistors in intel's older lake chips which we're expecting this fall these are intel's first desktop design with a mix of high performance and high efficiency cores and i am super curious to see how these perform we'll try to get our hands on one for an episode there's also a mysterious raptor lake chip in the mix but details are scarce next up intel's seven animator processors are being rebranded as four nanometer these chips will fully embrace euv which is the advanced manufacturing technique that we've covered before euv is currently only used in a few select layers in certain chips using it more should help streamline manufacturing because euv can help print certain layers in a single go things that used to require multiple passes and more time intel is claiming another 20 improvement in performance on this one and considering it should also come with a big improvement in density that seems plausible intel is hoping to get this node into gear in 2022 for 2023 products including their upcoming meteor lake chips meteor lake looks to be intel's first move into a fully chiplet or tile based design alder lake may have a mix of high and low powered processors but they're all built on one die kind of like apple's m1 chip while meteor lake could be built from a bunch of different tiles all linked together possibly with intel's stacked foveros chip design meteor lake or at least its zeon based granite rapids equivalent should also be the first wide scale use of intel's emib technology essentially when you're connecting a bunch of chiplets or tiles together there's a bunch of different ways to do it you can kind of just wire them all together through a circuit board which vastly oversimplifying infinity fabric is kind of how amd does it but this can add to latency or heat and power usage another way is to bond all the smaller chips that'll go into a processor onto a base of silicon called an interposer you can run more complex linkages through this and the interposer can even be active essentially acting as a network manager chip to help route power and data this improves performance and we see this a lot in high-end server chips like fpgas but as a big old slab of silicon the inner poser is expensive there's also the added challenge that you have to run the power circuitry through that silicon slab emib is intel's solution to this and it essentially uses just a small slice of an interposer as a bridge in between tiles this gets you some of the benefit of that high bandwidth active connection but it's cheaper uses less silicon and it stays out of the way of the power circuitry we actually first saw this in intel's strange katy lake chip from 2017 that paired an intel cpu with an amd gpu on one chip foveros is intel's solution for stacked processors first showing up in their lakefield chips the idea behind a stack chip is that by layering the logic the processing parts of the chip on top of the bits that do input output and networking and maybe even the memory you can reduce power consumption and improve latency by physically moving the bits closer to each other and also you can cram more chip into a smaller space we reviewed a lakefield device last year and while it was super efficient it was not terribly speedy and in fact those lakefield processors have already been retired foveros itself isn't going anywhere though intel has already announced two new versions of the design foveros omni allows for stacking chips of various different sizes this allows for some more flexibility in design and also for so-called cantilevered chips where the top layer is the largest this enables you to have so-called wiring columns around the edges where power delivery can reach directly to the motherboard below without having to pass through the other layers of the chip foveveros direct bonds the copper wiring directly together without the need for solder bumps this requires some pretty amazing precision and for both halves to be incredibly even and smooth but should allow for higher efficiency connections and also for more tightly packed contacts these ideas are all cool but we haven't really seen a high performance foveros design yet so we don't really know what to expect and at the high end these stacked chips may run into problems with heat dissipation still you can see a situation where emib and foveros might be used to design sort of a modular high efficiency cpu paired with a fast gpu and a bunch of high bandwidth memory up next intel's planned three nanometer node formerly seven plus seems a lot like four or maybe a 4 plus if you will it uses euv on even more layers and and some improvements in power delivery with a claimed 18 increase in performance we don't have a lot of other details here with production starting in 2023 probably for a 2024 launch intel makes it pretty clear through all of this they are targeting performance per watt as their main metric and this is a super important criteria improving performance per watt means the chip can either do more work for the same amount of energy or can perform the same job using less power the one caveat here is that intel's high-end chips currently run pretty darn hot with some of them drawing over 200 watts of power because what i'm trying to say here is that improving performance per watt doesn't necessarily mean you're getting a faster chip it could just mean a processor that can do the same work as today but for less energy either way i look at these stats and it does point to big improvements but with the real focus on performance for what i wonder if intel is a little less confident about reclaiming the absolute performance crown moving along the previous 5 nanometer node is now renamed 20a a is an angstrom for 110 billionth of a meter so two nanometers remember it's not clear if these names are actually measuring anything but a silicon atom is about two angstroms across so tiny this node will bring a transistor redesigned to ribbon fets we covered these in our ibm video that you can go check out but the idea is that at these very tiny sizes transistors get sort of leaky and hard to control by wrapping the gate material entirely around a silicon nanowire or ribbon you can more easily control the flow of power through the transistor and by varying the width of the ribbon's designers can actually tweak the electrical properties of the transistor for different parts of the chip these ribbon transistors will be paired with something called power via design this is intel's take on something called backside power delivery see in a typical chip the transistor layer is built first and then the wiring that controls information and power is built in progressively larger layers on top of that until it links into the contacts that connect to the motherboard the chip is essentially built upside down from how we see it the problem here is those two sets of wires have to snake around each other and can actually interfere the same way that high voltage power lines can disrupt cell phone signals the power connections can actually interfere with the wiring carrying data the power via design actually separates those two sets of wires in half and puts the logic in the middle this could enable say higher voltage wiring or higher frequency data connections 20a is planned to start manufacturing in 2024 with the next gen 18a in 2025 which will use high n a euv which is the next iteration of that manufacturing process and essentially uses a more powerful beam of uv light to create the image of the chip all of this is super cool and does point to cpus that could get much faster or more efficient now for all my doubts intel is going to need that improved efficiency because a lot of these designs sound pretty challenging from a thermal perspective packing more cores into a smaller space can generate more heat so can layering the chip components in a stack or burying the transistors in between two layers of connections intel's current chips suck down a lot of power and generate a lot of heat so i want to see how all of this works before i pass judgment the mixed core design of alderlake and the tile-based design of meteor lake are also unlike anything we've really seen from intel up until this point so it's hard to know what to expect i also just refuse to believe this timeline anything is possible but this cadence of releases bringing multiple new transistor designs to market very quickly seems unlikely if nothing else high in a euv has been delayed by asml currently the only company in the world that builds the high-end manufacturing tools that all of these chip makers need while that doesn't seem to have deterred intel asml is still only really capable at the moment of producing about 50 of these euv tools a year and each one can take months to ship install and calibrate still amd and apple have shown how a little competition can really spur an industry forward with those two i feel like we've seen that in the chip design space and now with tsmc and samsung becoming major competitors maybe that'll be the push intel needs to pull off something amazing let us know what you think how likely do you feel like that 2025 deadline is are you holding out for an alder lake or meteor like chip or are you waiting for a zen 4 or you're just gonna ditch pc altogether and get yourself an apple m2x or whatever they're gonna call that when we finally see it let us know in the comments below and be sure to like and subscribe and we'll catch you next time [Music] [Music] you

Channel: Engadget

Views: 95,730

Keywords: engadget, technology, consumer tech, gadgets, science, gear, tech

Id: FOFXO9HRZG4

Length: 13min 12sec (792 seconds)

Published: Wed Aug 04 2021