Comments Locked

155 Comments

Back to Article

  • THICCBOI - Friday, April 27, 2018 - link

    very interesting article, will be very interesting with Intel's cpu architecture and use of large monolithic dies will effect yields, AMD will have a clear advantage with Zen in the future due to there use of multiple smaller dies, they will be able to get much higher yields from a wafer with the same amount of defects compared to Intel. This will allow AMD to manufacture cpus on a smaller node for much cheaper compared to intel, forcing them into an impossible position as the nodes continue to get smaller
  • jjj - Friday, April 27, 2018 - link

    Intel is going to do chiplets too. AMD took the lead but Intel has plans too, you are already seeing their silicon bridge being used in the SKU with Vega and we should expect their own pseudo-discrete GPUs to use it it.
  • Spoelie - Saturday, April 28, 2018 - link

    Jim Keller was hired at Intel specifically to spearhead this approach.
  • Krysto - Saturday, April 28, 2018 - link

    Even if he lives up to the hype, you won't see the results of his work for at least another 3 years.
  • Krysto - Saturday, April 28, 2018 - link

    Also, he's not going to fix Intel's process node manufacturing issues, that's for sure. The best they can hope for is a cheaper to make architecture.
  • NeuralNexus - Friday, April 27, 2018 - link

    LMAO SO MUCH FOR INNOVATION...INTEL fanboys have been bragging about INTEL's manufacturing capabilities for so long. AMD on 7nm next year will really showcase how far AMD has come. Maybe Jim Keller can assist in ironing out the issues at INTEL.
  • Yojimbo - Friday, April 27, 2018 - link

    Will GF's 7 nm be any better than Intel's 10 nm for CPU manufacture? You can't just look at the nm numbers. Maybe it'll be worse. We also have to wait and see if GF makes their 7 nm targets.

    What does Jim Keller have to do with foundry operations?
  • NeuralNexus - Friday, April 27, 2018 - link

    Jim Keller going to INTEL might find a way to establish the next architecture so they have less difficulties manufacturing at a lower node to retain equal or better performance. As for GF, I do not know if their manufacturing process with be overall better. It will be a step in the right direction. GF usually license their node process from Samsung and improve upon it.
  • FullmetalTitan - Friday, April 27, 2018 - link

    GF only licensed the 14LPP process from Samsung due to issues with their own internal architecture development, subsequent designs were in house improvements. Their newer offerings are entirely in house.
    IIRC both Samsung and GF 7nm will be more dense than Intel 10nm, but even that metric is subject to some interpretation.
  • 0ldman79 - Friday, April 27, 2018 - link

    GF has a cross licensing agreement with IBM. I think the 7nm and 5nm are both based on IBM's work.

    https://hothardware.com/news/amd-7nm-finfet-vega-2...
  • caqde - Saturday, April 28, 2018 - link

    Not really cross licensing more like own. But it would be based on their work which is now technically their work I believe there was an agreement to manufacture PowerPC chips in this acquisition agreement though. https://www.globalfoundries.com/news-events/press-...
  • HStewart - Saturday, April 28, 2018 - link

    Intel 10nm is not out yet and just because Samsung and GF has a lower number - does not mean it is more dense.
  • Wilco1 - Sunday, April 29, 2018 - link

    Sure it does - density is what matters, and 7nm TSMC/Samsung/GF are all more dense than Intel.
  • Klimax - Sunday, April 29, 2018 - link

    I find lack of evidence amusing.
  • Wilco1 - Sunday, April 29, 2018 - link

    It's more amusing that you believe Intel marketing...
  • hecksagon - Monday, April 30, 2018 - link

    Process names are just good sounding numbers picked out by the marketing department. They have very little to do with actual measurements of average feature size.
  • peevee - Monday, April 30, 2018 - link

    ^^^^^
    This!

    Since ~"32nm" it is all fake.
  • Wilco1 - Monday, April 30, 2018 - link

    The number doesn't have anything to do with average feature size, and that has been true practically forever (20+ years). However it's not just marketing - the number does approximate relative density fairly well.

    So you can compare the numbers to get an idea of the "goodness" of a process. Of course one number is never enough to describe all of the (secret) details that make a process.
  • HStewart - Saturday, April 28, 2018 - link

    I don't believe that Jim Keller is about Zen, I think AMD people like to give AMD too much credit. Look at where he is aligned with SOC group and compare that too his experience with ARM and Apple device.

    I think Apple is key to everything Intel is doing here - along with Raju - to make new SOC's that please apple in there future.

    Not sure how reliable this is since it comes from WCCFTech - which often has new news before others have but reliability especially in this extremely crazy forums is questionable. But there was article about AMD need to switching away from GF 7nm to TSMC 7nm.
  • The Hardcard - Friday, April 27, 2018 - link

    Essentially about the same. There can be deep, esoteric debates about certain slight advantages here or there, but 2019 will be the first time in many years that Intel isn’t sitting a node if not close to two nodes ahead.
  • Wilco1 - Sunday, April 29, 2018 - link

    Today Intel is already 2 generations behind. Intel lost its advantage when 10nm came out.
  • Klimax - Sunday, April 29, 2018 - link

    You can assert that all you want, but lack of evidence is going to haunt you...
  • Wilco1 - Sunday, April 29, 2018 - link

    What lack of evidence? The fact that Intel is behind is indisputable. See for example: https://www.semiwiki.com/forum/content/6713-14nm-1...
  • hecksagon - Monday, April 30, 2018 - link

    Data is over a year old. A lot can happen in a year.
  • close - Monday, April 30, 2018 - link

    Yeah... like another 1 year delay. That will definitely put them in front.

    It's a well known fact that the more you delay, the farther in front of everyone else you'll get...
  • peevee - Monday, April 30, 2018 - link

    Intel lost the advantage 2 years ago with the failure to release 10nm. They are not ahead already, but not really behind yet (only in the fake marketing "node" numbers).

    AT, please write an article on comparative technologies, today. Transistor densities, current leaking, voltage/frequency curves...
  • Wilco1 - Monday, April 30, 2018 - link

    Intel has been behind since denser 10nm processes were released. Given 7nm is in production they are now 2 generations behind (Intels 14nm density is >3 times worse than TSMC 7nm). Next year the foundries will move to EUV which will bring further gains, so there is a chance for things to get worse.
  • peevee - Monday, April 30, 2018 - link

    Again you are comparing the fake "nm".
  • Wilco1 - Tuesday, May 1, 2018 - link

    No, I'm comparing real transistor density. See my link above. It's indisputable that Intel is 2 generations behind today, and won't be able to catch up even if they released 10nm today.
  • Santoval - Friday, April 27, 2018 - link

    GloFo/TSMC/Samsung's 7nm nodes are expected to have a higher transistor density than Intel's 10nm node, though not much higher. Intel has reported a 101 million transistors per mm^2 number, while the 7nm nodes of the others have been estimated to have transistor densities in the 115 to 135 million per mm^2 range.

    So, if (and that's a big "if") we assume that the density of the FEOL stack is the most important metric, the other three will still be ahead of Intel in 2019 and 2020, though not by very much. Of course transistor density refers only to quantity of transistors, not quality, and is unrelated to the BEOL stack. So it remains to be seen which node will perform best in terms of power, performance and efficiency.
  • Krysto - Saturday, April 28, 2018 - link

    A 15-35% increase in density sounds like a huge increase to me.
  • HollyDOL - Saturday, April 28, 2018 - link

    Ye, but it leaves too many unknowns... For example power-frequency-cooling factor... If for example higher density process forces lower final frequency duo to cooling issues, it might be in the end on par with Intel's 10nm.
    We won't know till both get benchmarked, until then we can just guesstimate and hope there will be some nice perf gains coming out of this fight.
  • Krysto - Saturday, April 28, 2018 - link

    Will Intel's 10nm ever be real?
  • Klimax - Sunday, April 29, 2018 - link

    Allegedly t is already shipping in smaller numbers, so likely around August/September we'll get to see some of notebook/tablet models with them.
  • peevee - Monday, April 30, 2018 - link

    Real also mean profitable. If the CPUs are too expensive (fixed costs are already sky-high, low yields mean high marginal costs too), customers will not buy them, and Intel will not produce them.
  • goatfajitas - Friday, April 27, 2018 - link

    You guys are funny (in a lame way) about your fanboy/non fanboy stuff... Most people, myself included just get what is best and/or best bang for the buck when they go to buy. At that AMD is looking pretty sweet these days.
  • Maxiking - Friday, April 27, 2018 - link

    Zen+ is on 14nm+ while being told by the marketing department it uses 12nm. So guess how will it be with 7nm?
  • Spunjji - Friday, April 27, 2018 - link

    That's Glofo's marketing department. They can call their process whatever they think best reflects its properties, and none of us on this site are under any illusions about what it actully entails...
  • hecksagon - Monday, April 30, 2018 - link

    AMD doesn't own any manufacturing capability so it will show nothing about how far AMD has come. Process size means entirely different things to the marketing departments at Samsung, TSMC, GF, and Intel, and do not reflect the size of any particular feature. Intel's 10nm is likely effectively the same as 7nm at another manufacturer, or 8nm at another.
  • mdriftmeyer - Monday, April 30, 2018 - link

    Chip design architecture and Ph.Ds in Physics don't share the same worlds. Keller has nothing to do with any of the fabrication technologies and research. He's going to help their teams redesign a SoC architecture, that is all.
  • evilpaul666 - Tuesday, May 1, 2018 - link

    Intel has had better process technologies for...forever. That's a simple fact. If you're arguing that you should probably check the mirror for "fanboys."
  • Achaios - Friday, April 27, 2018 - link

    Sounds like a fiasco.

    If, as the AMD fanbois say, AMD goes to 7nm next year and Intel struggle again to release their 10 nm (the latter is very, very likely IMO), that would place AMD at the cutting edge of technology, perhaps launching CPU's with faster single-threaded performance compared to Intel, something that hasn't happened since Northwood days.

    Very exciting and very cool for consumers, as the competition will likely bring huge discounts.
  • Santoval - Friday, April 27, 2018 - link

    AMD will almost certainly release Zen 2 based CPUs at 7nm in Q2 2019, either by GloFo or by TSMC or both. If by that time Intel has not released Ice Lake (I strongly doubt they will), they will fall behind in performance, both single and multi-threaded. The design of Zen 2 will surely address the original Zen's weaknesses against Intel CPUs : Lower IPC, lower clocks, lower overclocking potential, and a very high L3 and CCX to CCX latency.

    These were already partly addressed with Zen+, but there is only so much you can do without redesigning a CPU. Who knows, they might even do away with the CCX+CCX concept altogether, and keep the Infinity Fabric (a faster version of it) only for inter-die communication of Threadrippers and above. Or not, if they can speed up the CCX to CCX link significantly.
  • Flunk - Friday, April 27, 2018 - link

    I think the 2 CCX thing is a way to core harvest cores that only have one working CCX to use as lower end CPUs. Similar to how they design GPUs.
  • Spunjji - Friday, April 27, 2018 - link

    I suspect they'll have been working on the latter. A process improvement alongside a redesign would give a lot of room to drop latencies and/or increase clocks on their chip-to-chip bandwidth.
  • Krysto - Saturday, April 28, 2018 - link

    They won't do away with CCX, but they may use 8-cores per CCX. That would be great, but not sure if they'll do it at 7nm or 5nm.
  • willis936 - Friday, April 27, 2018 - link

    I wonder if Intel will feel the heat turned up in the next few years and return to their tried and true anticompetitive practices.
  • benedict - Friday, April 27, 2018 - link

    No, Intel is so deeply entrenched in the market, that they could make processors twice as slow as their current ones and still sell a ton and have 70% market share.
  • Krysto - Saturday, April 28, 2018 - link

    My guess is AMD will gain over 50% market share in PCs, probably over 30% in notebooks, and over 15% in servers, over the next 3 years.
  • Klimax - Sunday, April 29, 2018 - link

    That's so wildly optimistic that it can at best qualify as AMD Fanboy alt-universe wish.
  • close - Monday, April 30, 2018 - link

    AMD barely touched 50% market share in PCs (you probably mean desktops) at the peak of their popularity during Athlon 64 era - 2006. And the victory was short lived as 2006 was the year Intel launched Core and Core 2. AMD went down to 40% market share in one quarter.

    Intel knows how to play the game. So even when AMD has the upper hand and the better product Intel knows what to do to stunt their growth by any means necessary (including illegal ones). And they got away with it once so why stop?
  • willis936 - Saturday, April 28, 2018 - link

    Intel’s customers are businesses. Intel’s competitor is a drop in replacement. How are you going to convince a business to spend more on an inferior product? The only path I see is to implement anticompetitive practices to maintain their current customers. You see Nvidia doing it and they don’t even have a remotely threatening competitor.
  • t.s - Saturday, April 28, 2018 - link

    With lobbying of course, you silly!
  • t.s - Saturday, April 28, 2018 - link

    Agreed. Average Joe is so accustomed with Intel that they 1) never know that competitor exists or 2) still have a negative stigma over AMD, that they don't know how AMD become good competitor in this field now.
  • Yojimbo - Friday, April 27, 2018 - link

    "As it appears, right now Intel executives do not consider EUV technology ready for prime time in 2019, so the company’s engineers have to polish off the last DUV-only process (again) rather than jump straight to 7 nm."

    There are two factors preventing Intel from using EUV in 2019 if they wanted to, or even preventing them from accelerating EUV much more than their current plans in a time frame beyond 2019. Firstly, any such decision needs to be made far in advance. It currently takes ASML 18+ months to fill an EUV tool order. Secondly, ASML only can produce so many EUV machines at a time, owing to Carl Zeiss's ability to produce the optics, and they are already all filled up for 2018 and 2019. So one cannot infer from Intel's lack of EUV usage in 2019 what Intel's current view is on EUV's readiness for prime time, only what their view was 3 years ago, or whenever they were planning out their 10 nm node, and even then only what they though EUV would be ready for at the date they originally planned to be performing high volume 10 nm manufacturing.
  • p1esk - Friday, April 27, 2018 - link

    I'm surprised that this whole EUV ramp up takes such a long time. Given its seeming importance, and the untold billions invested in it, and number of years everyone has been working on it.
  • FullmetalTitan - Friday, April 27, 2018 - link

    The complexity of EUV is mind-boggling. There are issue creating the masks because they are many alternating layers. Pelicles are not close to ready and still attenuate too much beam power. Beam power itself was a major hangup, and just with the latest ASML platform hit a level deemed acceptable to foundries buying the tools. Throughput is an order of magnitude slower than DUV/immersion tools. The photoresists are a problem, since you are now using fewer photons with higher energy, shot noise becomes a large problem. Stochastic behaviors throw line edge placement for a loop at small geometries...
    This list is really a couple pages long but you get the idea.
  • Yojimbo - Saturday, April 28, 2018 - link

    No one bothered putting much effort into the pelicles or the masks until they solved the beam power problem because it could have been wasted effort. People have been confident for a while that the pelicle and mask issues are solvable and will be solved in the time frame necessary. In fact I think ASML says they have them now. At this point the attitude seems to be that EUV is pretty secure. TSMC and other customers (who, I don't know, but probably Intel, Samsung, and GlobalFoundries) have ordered second generation high-NA EUV tools for R&D purposes. The machines cost well over 100 million Euros each. High-NA EUV is not expected to be used until 2023 or 2025 (I forget which). So it seems clear that the foundries are pretty confident in EUV at this moment.

    From what ASML's CEO said, the reason the ramp takes a long time is because they need to gain experience with the production to assess where to make efficiency improvements. He said it was just the normal engineering learning curve for new manufacturing processes. That's why it takes so much longer at the moment to build an EUV tool than a DUV tool. The reason they can only make 20 EUV tools in 2018 and 30 in 2019 is because Carl Zeiss can only handle producing that many optics systems in those time frames, as they too are learning how to ramp up the EUV optics production. They are already at 0% margins for EUV systems in 2018, I believe. If they tried to push it and produce more they'd be losing money on the additional systems.
  • p1esk - Saturday, April 28, 2018 - link

    Will EUV technology they are developing now be useful or applicable for post-CMOS lithography? Like, I don't know, graphene or photonic circuits?
  • Yojimbo - Sunday, April 29, 2018 - link

    I don't know. I think it depends on whatever eventually works out, and what future discoveries are made that lead to that technology winning out. For some technologies they could potentially grow the circuits instead of draw them, in which case photolithography would no longer be the key technology. But, that would demand a major infrastructure shift. Volume manufacture using such a novel method surely is more than a decade out. At the moment, though, I don't think they even have a hopeful candidate to move ahead with. They are still doing basic research on various technologies. The next 10 years of progress will almost certainly rely on photolithography and right now EUV is the only economical way forward. I think there's a good chance, though, that many generations of process nodes will be operated at the same time going forward. Only the projects with fat margins will warrant the die shrinks since the decline in cost per transistor between generations is becoming rather small. So DUV and EUV will exist side-by-side over the next 10 years.
  • FunBunny2 - Sunday, April 29, 2018 - link

    "I think it depends on whatever eventually works out, and what future discoveries are made that lead to that technology winning out. For some technologies they could potentially grow the circuits instead of draw them, in which case photolithography would no longer be the key technology. "

    manufacturing tech isn't the issue. the engineers will always figure out some ways to make the widgets. the issue is at what feature size (and if that is one size fits all feature types, so let's just say minimum) does that feature go Heisenberg? no. quantum computing won't support deterministic problem spaces. until someone shows confirmatory math.
  • Yojimbo - Monday, April 30, 2018 - link

    Manufacturing technology is a big issue. Thousands and thousands of man-years has gone into understanding and refining photolithographic idioms. Billions of dollars have been spent on the photolithographic infrastructure. Whenever someone comes up with a new technology they are thinking about bringing to mass market a big question he must ask himself is "Can this be made with standard CMOS manufacturing techniques?" Of course if a technology shows enough promise, or if there is no suitable alternative, such a barrier can be overcome. But that's a far cry from there not being an issue there at all.
  • FunBunny2 - Monday, April 30, 2018 - link

    "Manufacturing technology is a big issue."

    at this very moment, that's true. even using known, and knowable, tech engineers could produce feature size at 2nm or 3nm, I suppose. but that's meaningless if such features go Heisenberg. the point is, of course, that Heisenberg is closer than it appears in the mirror.
  • peevee - Monday, April 30, 2018 - link

    You look like a person who knows something.
    Why the huge jump from 193nm to 13nm? What happened to intermediate wavelengths? Say, 100nm would suffer much (~60 times) less from shot noise compared to 13nm, but would eliminate good deal of multipatterning already.
  • Yojimbo - Tuesday, May 1, 2018 - link

    I don't know, sorry. I'd guess that the issue is one of materials. They need a light source, optics, masks, and resists for any wavelength. Perhaps they couldn't find any set of suitable candidates until they got down to EUV wavelengths. They were looking at X-rays, too, but in the end it was EUV that showed the most promise. I know that they were developing a 157 nm system. It was beset with problems, but I think they might have solved them. However, the industry settled on immersion lithography at 193 nm instead.
  • Yojimbo - Tuesday, May 1, 2018 - link

    Oh I found something. It's a a journal article by someone from Bell Labs of Lucent Technologies published in "Material Today" in 1999. It says:

    "Extension of conventional refractive optical lithography to wavelengths below the F2 excimer (157nm) is problematic due to absorption in the refractive elements. It is also very difficult to design and construct reflective optical systems with sufficiently large numerical apertures to allow printing at or below the wavelength of the illumination source. The idea of EUV lithography is to use small numerical aperture reflective optical systems at wavelengths much shorter than the circuit dimensions. The reflective elements for EUV use multilayer Bragg reflectors such as MoSi pairs to produce reflectivities up to nearly 70% at 14 nm. The mask in an EUV system is reflective and also uses the same type of multilayers. A plasma or synchrotron based source is used to illuminate the mask which is imaged by a system of mirrors onto the resist-coated wafer with a reduction factor of four. The optical systems require mirrors with unprecedented tolerances with respect to figure and finish.That is, the shape of the mirror must be correct in addition to the surface being smooth. The specifications are in the angstrom and in some cases, sub-angstrom range posing serious challenges for mirror fabrication, coating, and mounting."

    So, according to that, they have problems with refractive optics at wavelengths shorter than 157 nm and in order to use reflective optics they need much shorter wavelengths.
  • Spunjji - Friday, April 27, 2018 - link

    This x1000
  • Santoval - Friday, April 27, 2018 - link

    Jesus, so that means that Ice Lake is moved back to a Q1 2020 release at the earliest, probably Q2 for truly high volume. And Intel will have spent at least 5, perhaps even 6 whole years (2014 to 2019) releasing various 14nm, 14nm+, 14nm++, 14nm+++ etc CPUs.. If the 7nm yields of TSMC, GloFo (and Samsung) turn out to be much better Intel is going to be crushed next year by AMD, and ARM based CPUs will close even more the performance gap with Intel CPUs.
  • Kevin G - Saturday, April 28, 2018 - link

    Ice Lake is still looking like a 2019 launch, though probably later in the year. The design has taped out so the masks are done, just waiting on yields to improve. This was before Meltdown/Spectre. So even if
    Intel were to revise them, they'll still be waiting on the process to mature.

    What this does mean is that Cannon Lake is vapoware.
  • jrs77 - Friday, April 27, 2018 - link

    I guess it was only a question of time before we hit the wall with the process. Sure there are products made in 7nm allready, but they're nowhere near as complex. So I don't think that intel will bypass the 10nm process and go straight to 7nm, as we've hit the current limit imho, as to how small we can go with the current materials for highend CPU manufacturing.
  • Cooe - Friday, April 27, 2018 - link

    "Sure there are products made in 7nm allready, but they're nowhere near as complex."
    What the hell are you talking about??? Flagship GPU dies (Vega 20 in this case) are about as big & complex as chip designs get. All the major 7nm parts so far have been big Kahuna's.
  • jrs77 - Friday, April 27, 2018 - link

    They're not made in 7nm. Look at the rasterelectron microscope pictures of the so called 7nm processes... they're 14nm in reality.
  • Wilco1 - Friday, April 27, 2018 - link

    Sure the features of any process as never as small as the number suggests. That's been true at least since 130nm for all processes.

    However the 7nm processes from TSMC/Samsung/GF are definitely more complex and dense than Intel 10nm.
  • name99 - Friday, April 27, 2018 - link

    Or maybe not. 😀
    To be fair, no one has ever even seen an Intel 10nm chip in the wild, let alone cut one open. For all we know they are full of pm sized transistors and magic fairy dust.
  • Spunjji - Friday, April 27, 2018 - link

    Current Vega isn't but the test dies they have are...
  • Krysto - Saturday, April 28, 2018 - link

    Intel doesn't even plan to use EUV until 5nm, which given Intel's schedule so far, may happen in 8-10 years. Intel is going to get crushed in the process department over the next few years.
  • lefty2 - Friday, April 27, 2018 - link

    Great article. Another thing is that 10nm won't clock as well as 14nm++, so even when 10nm gets good yields it's not going to be suitable for desktop CPUs. (Basically, Broadwell over again)
  • GeorgeH - Friday, April 27, 2018 - link

    A thousand times this. It isn't enough for Intel to get 10nm working, they also have to make it competitive with their 14nm process. We are no longer in the realm (and haven't been for awhile) where making transistors smaller obviously translates into better CPUs.

    Had Intel spent years refining 22nm, there's every possibility that a hypothetical 22nm+++ node CPU could be more than competitive in many metrics with a current 14nm CPU. As Intel continues to make their 14nm process better, they're also raising the bar that their 10nm process must meet and making things more difficult for themselves.
  • haukionkannel - Friday, April 27, 2018 - link

    10nm allows more cores in smaller space! The clock speeds can go down because you have more cores to do stuff!
    Ofcourse games belongs not to that croup of programs, but games Are not important in the big picture...
  • peevee - Monday, April 30, 2018 - link

    Space is not the issue on desktop for a long long time.
  • Spunjji - Friday, April 27, 2018 - link

    They've only just exceeded the clock potential of Sandy on 32nm. Imagine 32nm+++... XD
  • ಬುಲ್ವಿಂಕಲ್ ಜೆ ಮೂಸ್ - Saturday, April 28, 2018 - link

    Not just clocks...
    We are not required to use Meltdown/Spectre mitigations with Sandy either

    Running known software securely with non sensitive data is fine for Sandy

    For sensitive data, low clocks and many cores "might" be acceptable with meltdown hardware mitigations @ 10nm

    However...
    History has shown that security will be thoroughly hosed regardless of what mitigations are in place

    Many low clocked Atom cores are simply not acceptable!
  • eddman - Saturday, April 28, 2018 - link

    Keep posting out-of-topic, troll comments, BJM.
  • Krysto - Saturday, April 28, 2018 - link

    Oh, the first-gen 10nm Intel chips are going to suck for sure, even more than Broadwell did. That's a given. You'd be a fool to buy Intel's Cannonlake after failing so hard to ship it. You can imagine that they have strong incentives to be misleading about how good it really is because they can't wait to just ship it already and be done with it.
  • Stochastic - Friday, April 27, 2018 - link

    This makes me think twice about holding out for an upgrade.
  • Krysto - Saturday, April 28, 2018 - link

    In 2019, the best upgrade will be to an 8c/16t Zen 2 chip anyways. I don't think there will even be a contest.
  • Klimax - Sunday, April 29, 2018 - link

    Are you paid by AMD marketing or what? How the hell you can even assert that? (Only AMD has samples of Zen 2 to know its performance characteristics) Or do you have access to full production Zen 2 chip?
  • eva02langley - Monday, April 30, 2018 - link

    You look like the one trying to promote Intel... just saying...
  • mpbello - Friday, April 27, 2018 - link

    TSMC announced 7nm HVM this month, Apple's 7nm A12 will be shipped later this year and will show everyone why it is time for Apple to kiss Intel goodbye. That move from Apple will open the floodgate and will be the start of Intel's market decline. (From a technical PoV, Intel's decline became obvious back in 2014 as they struggled to ship 14nm, but it will take that long for their technical decline to finally translate to market decline).

    A broader theme is the rise of Asian players in high technology industries, what is happening to Intel is happening to the US in a broader sense.
  • eva02langley - Monday, April 30, 2018 - link

    That's an interesting argument. I also believe Intel is not in a good position and I don't believe monopoly tactics will work ever again.
  • FunBunny2 - Monday, April 30, 2018 - link

    "I don't believe monopoly tactics will work ever again."

    he says as Trumpism marches us rapidly to oligarchs.
  • jjj - Friday, April 27, 2018 - link

    Intel pulls a Tesla, except their CEO doesn't sleep on the fab's floor.
    10nm was announced at IDC 2013 and targeted for 2015.
    Shipping low volumes means sampling, in Intel's proprietary language.

    This is big news for AMD, gives them a lot of extra room to gain share in all areas. Just this new delay could easily mean 3-4 extra billion in revenue for AMD in the next 5 years.
  • akamateau - Friday, April 27, 2018 - link

    Hmmm...

    Will AnandTech actually Benchmark 7nm AMD Ryzen and EPYC CPU's vs 14nm Intel CPU's?

    I expect that the on-line media WILL NEVER show Intel in a bad light.
  • MrSpadge - Friday, April 27, 2018 - link

    > I expect that the on-line media WILL NEVER show Intel in a bad light.

    Too young to remember the Pentium 4 days?
  • HStewart - Saturday, April 28, 2018 - link

    Don't worry about that forums of tech sites - do a very good job on doing that - but do seriously real customers care about that.
  • Ninjawithagun - Friday, April 27, 2018 - link

    Funny how this is 'okay' and not really a big deal for Intel to announce such a delay. Had it been AMD, the virtual CPU sky would be falling as we speak. Perspective...sponsored and paid for by Intel.
  • Maxiking - Friday, April 27, 2018 - link

    Damn, in every article about Intel so many amd butthurts, that HPET still hurts.
  • eva02langley - Monday, April 30, 2018 - link

    HPET is just showing something is damn wrong with Intel architecture. It is not affecting AMD at all. Also, we don't even know if this is an issue with just old games.

    By the way, enjoy your 1080p @ 144Hz bench for < 3% of newer built.
  • MrSpadge - Friday, April 27, 2018 - link

    Dude, stop these stupid accusations. Did you notice that you're typing this on the website which is accused of producing the best (=unfair) Ryzen 2 results?
  • Hifihedgehog - Friday, April 27, 2018 - link

    Meanwhile...
    “AMD Sampling 7nm Zen 2 CPUs Later This Year, Launching in 2019“

    https://wccftech.com/amd-sampling-7nm-zen-2-cpus-l...
  • HStewart - Friday, April 27, 2018 - link

    One extremely important thing that most people don't see - is that GF's 7nm does not mean it is better than Intel 10nm - just because the number. There is a lot more to process than just the number. Maybe the reason they are having lower yields is because the process is significantly more dense than any current or future process from other companies.

    It also possible that the recent threats of AMD with Ryzen and Windows for ARM ( more significant threat to Intel ) are not as big as some one think.
  • Maxiking - Friday, April 27, 2018 - link

    It is not about density in general, when Intel releases another process generation, it means they follow the Moore's law. Unlike GLO, TSMC, Samsung and their "nm" spamming.. for example Zen+, it was supposed to be on 14nm+ but they did a marketing stunt a few months ago and decided it to call it 12nm instead while using the same density as on 14nm.

    Then to hide it somehow, they compare it to 16nm Finfet and say they managed 15% density improvement and 10% performance gain. And everybody quoted them and was comparing it to the old 14nm. Even here.

    https://i.imgur.com/cQJnTPS.png

    Now this marketing works again, they say 7nm why it is highly questionable but as you can see, people buy it, they see 7, 7 is lower than 10 and it works. Then let's look at Intel's 14nm and the 14nm AMD use with Zen+. Intel's one is more dense, can clock up 5.2ghz. AMD ones? Less dense, can barely reach 4.2 on all cores then it melts..

    So yeah, hurray, 7nm.
  • Wilco1 - Friday, April 27, 2018 - link

    There is absolutely no doubt that 12nm is significantly denser than 14nm - did you see the Ryzen 2 die photo? That's a clear density gain unlike Intel 14nm+/++ which are lower density.
  • Maxiking - Saturday, April 28, 2018 - link

    https://www.semiwiki.com/forum/content/6713-14nm-1...

    "12nm" offers up to 15% density over 16nm Finfet. Do the math.
  • Wilco1 - Saturday, April 28, 2018 - link

    There is no mention of 12nm in that article. 12LP was announced by GF as a "15% density improvement over competing 16/14nm processes". That means 16nm TSMC, 14nm GF/Samsung, so that obviously implies 15% over 14nm GF too since GF and Samsung use the same process. How it's done hasn't been detailed, but it may be a combination of BOL scaling and reduced track libraries.

    Zen+ however didn't take advantage of that 15% density - both Zen and Zen+ use the same 9 track libraries as explained here:

    https://www.anandtech.com/show/12625/amd-second-ge...
  • Maxiking - Saturday, April 28, 2018 - link

    No, it doesn't, because 14nm GLO and 16nm TSMC do not use same density, so that's why they used up to.

    That 15% is fabricated number to make things look better. 16nm density is 28.2, 14nm is 32.5, do you see that big difference and how much you increase the number when you compare it to 16nm and you up to? Comparing it to 14nm, the improvement number would be way, way, way lower.

    Intel's 14nm is up to 1.5 denser than the competition of GLOFO so even with those up to 15%, 12nm doesn't stand a chance.
  • Wilco1 - Saturday, April 28, 2018 - link

    The fact is that 12nm gives a significant density gain over GF 14nm, there is no way you can claim it is zero.

    Intel's original 14nm transistor density is 37.5 mt/mm^2 vs GF 14nm's 32.5, which is... about 15%. Note Intel's 14+/++ are at least 10% less dense, so it's pretty obvious GF 12nm will be denser.
  • ilt24 - Monday, April 30, 2018 - link

    Who's 14nm are you talking about? During the Q1 earnings conference call someone asked if the GF 12nm process was a die shrink or just performance improvements, Lisa said "It is not a die shrink."
  • HStewart - Saturday, April 28, 2018 - link

    Yes but some people are naïve and just believe that 7nm from one manufacture is more dense than 14nm from another. Especially understand that there is more to the process than number - but marketing wants customers to be believes it just about the number. No wonder AMD head marketing left the company - some time reality hurts.

    Also there is big difference in manufacturing technology - density is not always flat - it can be 3D.

    I believe Intel has something super impressive with their 10nm and things like that take time. They are higher high up to help re-inforce this not replace it technology.

    Of course like everything else on the internet, this is just my opinion.
  • Wilco1 - Saturday, April 28, 2018 - link

    You've got to be beyond naïve to believe that Intel 14nm is somehow better than 7nm from the foundries. Do you seriously believe that?
  • HStewart - Sunday, April 29, 2018 - link

    What I heard is that 7nm competition is close to what Intel would have is 12nm. I basically heard it will be denser than intel 14nm - but not as much as the Intel 10nm - I am comparing next generation technology.
  • Wilco1 - Sunday, April 29, 2018 - link

    No you must be thinking about 10nm from the foundries which shipped last year. The current generation 7nm is 15-30% denser than Intel's upcoming 10nm. However when 10nm ships next year it will compete with 7+ EUV processes which are even denser.
  • Wilco1 - Friday, April 27, 2018 - link

    No that's completely false. TSMC/Samsung/GF's processes are more dense than Intel 10nm. That's just the raw logic and SRAM density - generally in actual designs the difference is even larger. Several 20nm designs were shown to have higher density than Intel 14nm - and that was at a time where Intel was ahead.
  • Spunjji - Friday, April 27, 2018 - link

    That was a big chunk of the problems with 20nm... too small a feature size for planar transistors. 14nm was less dense in some circumstances but used that well for overall characteristcs.
  • ijdat - Saturday, April 28, 2018 - link

    TSMC/GF/Samsung 7nm processes are actually very similar (MMP, CPP) to Intel's 10nm; gate density is if you use the standard high-density low-power 6T libraries, but there are also 7.5T and 9T libraries similar to Intel's which which have lower density, higher speed and higher power to get similar results to Intel.

    The real yield problem with Intel 10nm could be multiple patterning (which everybody does -- but maybe they pushed it a bit too far) or the fact that they are the first to use cobalt instead of copper for a lot of low-level interconnect and vias, or a combination of both. Cobalt gives lower resistance, but as a completely new material could also be the underlying cause of their yield problems -- and having designed it in, they can't easily go back to copper.

    The foundries took the safe option and stayed with copper (maybe with cobalt for a very few layers and/or vias?) and slightly less aggressive multi-patterning, and don't have the same yield problem as Intel (yes I've seen internal foundry data for 7nm).
  • SaturnusDK - Saturday, April 28, 2018 - link

    No. What Intel planned originally was for their 10nm to be similar to about 7.7nm from other foundries. They've scaled it back twice now and are closer to about 8.6nm. That's a massive loss in expected density they've had to make just to get the process working.
  • HStewart - Saturday, April 28, 2018 - link

    Don't tell this to AMD fans - they will not believe it - they think is not as good and believe only in nm numbers - reality will come when Intel releases to the 10nm process and people again will call Intel a monopoly because nothing will be close to it.

    Just gluing 2 or 4 8 core - lower power Zen cores together - does not make it better.
  • Wilco1 - Saturday, April 28, 2018 - link

    Intel's 10nm process is already behind even if it was released today, so in a year it will be even more behind. The competition will make EUV chips next year. Hyperscaling done right!
  • HStewart - Sunday, April 29, 2018 - link

    Well it takes time to get real technology improvement do right - not just gluing 2 or 4 8 core zens together. How can you say "Hyperscaling done right" when AMD people are dropping like flies.
  • Wilco1 - Sunday, April 29, 2018 - link

    Actually combing multiple smaller dies is a great idea. It's standard in mobile SoCs. We will absolutely see much more of this in the future.

    Btw aren't you always promoting the multiple dies in your laptop? Or do you think it is only a good idea when Intel does it?
  • peevee - Monday, April 30, 2018 - link

    "Cobalt gives lower resistance"

    How? It's resistivity is 3 times that of copper.
    http://www.radio-electronics.com/info/formulae/res...
  • Wilco1 - Monday, April 30, 2018 - link

    Copper needs a fairly thick barrier layer around it as it otherwise poisons the silicon. Apparently that barrier layer doesn't shrink much, so there is less and less copper. With cobalt the barrier can be made much thinner, reducing resistance. https://en.wikipedia.org/wiki/Copper_interconnect
  • NEGuy123 - Friday, April 27, 2018 - link

    What I find interesting is... I don't see much mentioned about AMDs huge potential they have with their INFINITY FABRIC.

    You have heard AMD mention their Starship chip. I have heard this chip mentioned to have a 48 core & 64 core variant.

    I have see other reports of AMDs move to 7nm having 2 dies. A 12 core, and 16 core die.
    If AMD can actually pull off this core count... I don't see how Intel can stand a chance... for a while

    All this can equate to...

    AM4 socket can get a 12 core / 24 thread CPU
    Threadripper can get a 24 core / 48 thread CPU

    EPYC can get a 48 core / 96 thread CPU
    or EPYC can get a 64 core / 128 thread CPU.

    I personally think people do not grasp the magnitude of what AMD has planned!!! And I doubt Intel ever imagined AMD would or even COULD do this.

    I am super excited to see the 7nm from AMD and the possibilites it will bring... with Infinity Fabric.
  • Krysto - Saturday, April 28, 2018 - link

    I think it would be amazing if AMD switched to 8-core CCX on 7nm. But they probably won't until they get all of their profits in check and get the GPU division strong, too. So they may wait at least until Zen 3 or 5nm to do that.

    But yeah, I think 2019-2021 will be AMD's years for sure, especially with the new consoles launching, too. AMD stock will be $50 by 2021, I think (I don't own any myself but I wish I did).
  • eddman - Saturday, April 28, 2018 - link

    Yes, and intel have EMIB (which got its first application with kabylake-G) so they do have something to counter with. The only question is, when will it materialize?
  • eddman - Saturday, April 28, 2018 - link

    By materialize I meant "When we'll see multi-die EMIB CPUs from intel?".
  • HStewart - Saturday, April 28, 2018 - link

    Even though I am typing this on a Kabylake-G , I am not sure this was first adaption of the technology - a lot of PGLA stuff - and i believe the real application will come in next year of two with Raju based changes for Discrete GPU's.
  • twtech - Friday, April 27, 2018 - link

    This is not surprising. Any time a company puts focus in their hiring efforts on anything but competency, less than competent results are likely to follow.
  • peevee - Monday, April 30, 2018 - link

    Hear hear.
  • nobodyblog - Saturday, April 28, 2018 - link

    I suspect intel will change design of cannon lake or ice lake or does anything around its 10nm, people are crazy for these statements. Jim is probably a manager, not a good designer web kids claim.. And in CPU, intel is the best of all....... What the heck they say??

    Thanks!
  • Anymoore - Saturday, April 28, 2018 - link

    Not sure how they got hexa-patterning. The IEDM paper said they used quad-patterning, self-aligned type, on three layers, at most. TSMC has been using this for 10nm FinFETs already. There does not appear to be any excessive multi-patterning. However, Intel did use cobalt in combination with the most difficult multi-patterning. I wonder if that aggravated the situation. Adding EUV would only make things more complicated with two new process additions (EUV and cobalt).
  • HStewart - Saturday, April 28, 2018 - link

    The following job offering may provide some insight in what Intel is doing with next generation of CPU's - some call Oceans Cove

    https://jobs.intel.com/ShowJob/Id/1597785/Senior-C...

    I don't believe these are the Core CPU group - in the early 90's I met someone on the P5 performance group. But that did not work - it was odd - I had 7 years of x86 programming experience on OS level and they want C/C++ programmers.
  • HStewart - Saturday, April 28, 2018 - link

    I don't believe Jim Keller is actually in the core - core CPU group - but that is hard to tell with Intel.
  • Yojimbo - Sunday, April 29, 2018 - link

    Yeah, I read somewhere that he is going to be working on SoCs for Intel. I don't remember the source for that information, though. In any case, if Keller were taking over the lead of their Xeon/Core architecture design that would be a major shift that Intel would likely come out and announce to their investors.

    It's possible that Keller will be working on Intel's self-driving car efforts similar to what he was just involved with at Tesla. Intel has various technologies they want to mold together into a product, including Atom and what they got from Movidius, Altera, Nervana, and Mobileye.
  • HStewart - Sunday, April 29, 2018 - link

    Just like Raja, this does not mean the technology in SoC's would not end up in the consumer line - as technology gets more advance - more parts of CPU get integrated in the chip. One will be foolish to think it is chance that Raju change and development of 8705G - but AMD graphics is only temporary - I would also think it chance that both Jim Keller and Raju have relations with Apple.

    I also believe that idea of Apple leaving Intel is false information.
  • dhjffryhhjfgbj - Sunday, April 29, 2018 - link

    Intel (tmg) does not have a technology problem, but a culture problem. The motto is "everyone is expendable, for any reason I (the manager) see fit". Intel (tmg) is currently losing engineers and technician s faster then the technology cycle (if a tech needs 4 years, engineers leave in <2), the system has purged itself of any competence, and rewarded the worst of the worst. You can't do development in such a toxic and regressive environment! Let's rephrase:"you can't use third world management skills to lead a high tech development company". Intel will need years to recover from the damage done in the last six-ish years.
  • peevee - Monday, April 30, 2018 - link

    Can you elaborate? Are you talking about Sohail Ahmed?
  • dhjffryhhjfgbj - Saturday, May 5, 2018 - link

    Sohail's iron fist obliterated anyone who dared to oppose him...he obliterated out entire generations of engineers from intel within a year or two. TMG is now nothing more then a high-tech kinder garden, where people with 2 years experience are considered as senior. He is a genius but his fanatic perfectionist attitude created an unworkable environment. 10nm is not a technological problem in intel but organizational. The system is Sohail, Sohail is the system.
  • eastcoast_pete - Sunday, April 29, 2018 - link

    Whichever way Intel will spin this, they got caught napping. Intel's supremacy in laptops, desktops and servers was based on being ahead in both uarch and lithography , and both are under attack. Samsung and TSMC have both invested multi-billions $$$ in new fans and extreme ultraviolet technology, haven't heard similar from Intel. Really hope their recent exit from wearables means a re-focus on their core strength. In the meantime, you go, AMD! A strong competition is what makes companies push boundaries. Zen and Epyc plus the move to 10nm is just the kind of fire under chipzilla's butt that gets them going again.
  • eastcoast_pete - Sunday, April 29, 2018 - link

    strike fans, I meant fabs - damn auto correct
  • HStewart - Sunday, April 29, 2018 - link

    Actually issue is Intel Wanna Bee's - internally to Intel - everything is going good - actually hiring new people and coming up with new technology. Of course they realize some technology like wearables were false investments and they decided to get out of completely.

    You can tell these false claims - by simply what they call Intel "chipzilla" which is sure a dead giveaway of false mis-information.
  • Wilco1 - Sunday, April 29, 2018 - link

    Intel has been closing various departments recently, large markets were abandonned, many products were cancelled. And the now 4 year delay of 10nm must hurt.

    And you're seriously claiming all is going well as planned??? What a genious plan that is!

    So tell me, how do we know you aren't the one who is spreading all the misinformation?
  • Arbie - Sunday, April 29, 2018 - link

    "they call Intel "chipzilla" which is sure a dead giveaway of false mis-information"

    Chipzilla is used everywhere. It's not pejorative.

    "False mis-information" - is that true information?
  • zodiacfml - Monday, April 30, 2018 - link

    Crazy. I thought this will be the last generation of the 14nm products.
    They better lose the time and burn cash with EUV equipment or use it for 10nm products as an initial phase of EUV manufacturing.
  • peevee - Monday, April 30, 2018 - link

    Was somebody at least fired for the massive failure?
  • jardows2 - Monday, April 30, 2018 - link

    depends on the gender, ethnic, and sexual identity of the person responsible.
  • peevee - Monday, April 30, 2018 - link

    "This not only lengthens Intel’s manufacturing cycle (which by definition rises costs) and the number of masks it uses, but also has an effect on yields."

    And the only point of decreasing size of the features is to decrease cost per transistor. Which is increasing now. End of Si, finally?
    Given that other companies' 7nm is ~ Intel's 10nm, can we expect the same fiasco?
  • FunBunny2 - Monday, April 30, 2018 - link

    "End of Si, finally?"

    only if there's an alternative element:
    1 - equally abundant
    2 - equally cheap to extract
    3 - equally efficient to use

    the real problem is the decreasing number of atoms/feature. doesn't matter which element is used for construction. once you've crossed into Heisenberg territory, there be dragons there.
  • peevee - Monday, April 30, 2018 - link

    "1 - equally abundant
    2 - equally cheap to extract"

    Price of sand is non-issue in all chips, even in memory and flash, let alone CPUs and GPUs.

    Carbon is cheap and abundant, but even expensive Ge might be economical if it saves size (by raising frequencies) or manufacturing costs.

    Also, by "the end of Si" I meant the end of manufacturing improvements on Si. And it is not a physical barrier, it is an economic barrier. If you don't decrease cost per MIPS, decreasing feature size make no sense for mass market. And 10nm (Intel) bangs against that barrier for 5 years now with no win in sight, just like non-EUV "7nm" from other fabs will be with no doubt.
  • FunBunny2 - Monday, April 30, 2018 - link

    "but even expensive Ge might be economical if it saves size (by raising frequencies) or manufacturing costs."

    without proof that Ge can be used to produce features smaller, and still deterministic, than can be done with Si, there's no reason to believe the end product would be cheaper.

    "Germanium ranks near fiftieth in relative abundance of the elements in the Earth's crust. "
    from the wiki.
    "Over 90% of the Earth's crust is composed of silicate minerals, making silicon the second most abundant element in the Earth's crust (about 28% by mass) after oxygen."
    from the wiki.
  • peevee - Monday, April 30, 2018 - link

    "without proof that Ge can be used to produce features smaller, and still deterministic, than can be done with Si"

    Smaller features is not the goal in itself. 10GHz CPU can be half the size (in transistors) of 5GHz CPU and still win on performance. Just an example. Of course it is pure question whether or not Ge is able of that. I am NOT saying that it is. Just that the industry have to look at something else very hard at this point. Even if EUV works eventually.
    It is quite possible that at some point we will start seeing hybrids, like Si logic on graphene buses on separate layers.
  • FunBunny2 - Wednesday, May 2, 2018 - link

    "10GHz CPU can be half the size (in transistors) of 5GHz CPU and still win on performance."

    yes and no. that requires a boolean (or some other) logic that can build an ALU, etc. with 50% of transistors. that's an improvement which doesn't depend on the atom used. if you really mean chip real estate, Ge is more than twice the size of Si, so said ALU would have to be 25% of the Si equivalent. since Ge is so much fatter than Si, one might speculate that a feature of a given atom count could handle more current/power.

    yes, Ge/Si hybrid gates are being explored. whether they're a slam dunk way to avoid Heisenberg???
  • xeroshadow - Tuesday, May 1, 2018 - link

    And why wouldn't they? Despite AMD creating competition again, actually making profit for a whole year and great reviews from tech sites, Intel still made more profit in one quarter than AMD all year. When AMD has a billion in profit in one quarter then AMD fans can celebrate. Until then, Intel still gets to coast on its market dominance, superior manufacturing(at least in quantity that AMD can't match) and I'm sure the many contracts they have with the big OEMs. That is something that will take AMD years to claw back because even when they have a competitive product, it seems some big players stay away due to lack of manufacturing in quantity.

    Despite glowing reviews on AMD products they still always include a "but Intel has" line that always makes the reader second guess. They could coast on the i7 8700 and other Coffee Lake processors for another couple years. Unless AMD comes out with something that costs less, has 20% better performance AND uses less power than a competing Intel product, then Intel has no need to rush 10nm.

    Nearly the same goes for Nvidia not including back door dealings like GPP.
  • peevee - Tuesday, May 1, 2018 - link

    "superior manufacturing(at least in quantity that AMD can't match)"

    AMD does not have manufacturing at all, but can manufacture at fabs more than market can consume.

    But the reality is most CPUs today require built-in graphics (because they are used by businesses or normal home users, not extreme gamers), and Intel offers better CPUs (for example, 6 cores vs 4 at the top).
  • Boston Jackson - Monday, May 7, 2018 - link

    The first rumors about Intel’s problems with the tech started to spread in early 2015 as the company delayed the installation of equipment needed for its 10 nm manufacturing process.

    Showbox for iPad, iPad 2, iPad(3rd generation), iPad (4th generation), iPad Air, iPad Air 2, iPad Mini (1st generation), iPad Mini 2, iPad Mini 3, iPad Mini 4, iPad Pro, etc.

    https://showboxbuzz.com/download-showbox-for-iphon...

Log in

Don't have an account? Sign up now