Comments Locked

68 Comments

Back to Article

  • NickConrad - Thursday, September 30, 2021 - link

    Intel 4 is a 7nm node, not a 4nm node, if I'm not mistaken. Makes the title of this piece a bit of a head scratcher for me. Typo? My own dumb dumb ignorance?
  • Ian Cutress - Thursday, September 30, 2021 - link

    What do you mean '7nm node'? I'm not sure that has a definition, given that the names are meaningless, and have been a while. Intel 4 is designed to compete with TSMC N5 on its high-density libraries.

    But also, 'Intel 4' is a grammatically difficult phrase. Consider writing

    'Intel is using its Intel 4 process' rather than 'Using Intel's 4nm process'

    Just makes it a headache to write. So for the sake of ease of use, you're likely to see me use 4nm in titles. If they'd have called it N4 or 4N or 4F or something, it'd be so much easier to write about it.
  • NickConrad - Thursday, September 30, 2021 - link

    https://images.anandtech.com/doci/16823/Intel%20Ac...

    According to this slide from a previous post on your site, Intel 4 was previously known as 7nm. My interpretation of that is that Intel renamed a 7nm node as opposed to making a 4nm one.
  • drothgery - Thursday, September 30, 2021 - link

    Despite seeming to be based on measurements, process names (even at Intel) have been purely marketing at least since they started naming them in nanometers rather than microns. Yes, the smallest features on Intel 4 are over 7nm, not around 4nm, but that's true of TSMC N5 too. If it wasn't we'd be a lot closer to the end of Moore's law on planar silicon than we are; as far as I know, sub-4nm features really aren't physically possible.
  • GeoffreyA - Thursday, September 30, 2021 - link

    This whole nm nonsense of the industry ought to be done away with. Bunk is bunk, whether it comes from Intel or TSMC.
  • m53 - Thursday, September 30, 2021 - link

    Intel wanted to do away with it by coming up with an industry standard based on density and performance. But TSMC and Samsung didn't want to. Hence Intel had no other option but to rename theirs to match TSMC naming.
  • Alistair - Thursday, September 30, 2021 - link

    density changes depending on what you are making, not just the node, as this chip for example is only 1/3rd as dense as we are expecting for Intel 4... so no, your statement that Intel was the good guy here and came up with an alternative is false (like somehow they wouldn't play along)
  • m53 - Friday, October 1, 2021 - link

    Density changes on what you are making. NAND cell density is very different than logic density. Hence, you may not know that intel proposed a weighted average of both types using the following formula:

    0.6 x NAND2 Tr Count / NAND2 Cell Area + 0.4 x Scan Flip Flop Tr Count /Scans Flip Flop Cell Area = # Transistors/mm2

    If this pre-production Intel 4 process has lower density than the actual healthy HVM ready Intel 4 process, then why should both of these processes have the same name? Shouldn’t the production-ready higher density process have a different name that people can easily understand and differentiate? Should we just mention the weighted average density instead of the “nm” which doesn’t mean anything?

    So yes, that proposal, even with it’s flaws, was far better than nm based naming which is basically meaningless.
  • Spunjji - Friday, October 1, 2021 - link

    @m53 - agreed here, in that density-based naming would at least tell us *something* meaningful about the process, though not everything.
  • Spunjji - Friday, October 1, 2021 - link

    @m53 - That would be sad if true, but it's not exactly accurate. Intel dropped the attempt some time between relaxing the density of 14nm for better performance and failing to hit their projected density for 10nm, then relaxing that for 10+.

    This isn't a specific commentary on their choice of rename here, just to point out that Intel have at least as much of an investment in concealing the true functional density of their processes as the other two major players.
  • Silver5urfer - Thursday, September 30, 2021 - link

    Thank you for that, I would have been lost thinking how come Intel managed to keep up with the lithography lead loss. It's insane how Intel just pushes this hardcore tons of Lakes and now this random marketing term.
  • BedfordTim - Thursday, September 30, 2021 - link

    Two figures down the page in the same article, there is a table comparing transistor densities for various foundry's nodes. Intel 4 while previously called 7nm is equivalent to TSMC's 4/5nm. As Intel have changed to using something very similar to something TSMC's nm it is not unreasonable to use TSMC's figure.
  • shabby - Thursday, September 30, 2021 - link

    Blame intel not yourself, it should be labelled as intel 4. The more you write it as intel 4nm people will actually think it is 4nm.
  • m53 - Thursday, September 30, 2021 - link

    People commonly label TSMC N7 as as 7nm. I don't see people complain there. What's wrong with calling Intel 4 as 4nm?
  • shabby - Thursday, September 30, 2021 - link

    Because that's what intel calls it.
  • Ian Cutress - Thursday, September 30, 2021 - link

    If you write as much stuff as I do, having to continually say 'Intel is using Intel 4 process' is basically doubling up the word Intel every time you write it. The title would have literally been:

    'Intel's first Intel 4 EUV Chip'

    I decided to just bite the bullet and in titles at least, call it 4nm. In the same way people write TSMC N7 as TSMC 7nm. At least in the text I don't have an appreciable word limit and can word my way around a really grammatically awkward phrase. If they'd have called it Intel N4, or 4N, or 4F, or 4FQCS, then that'd be easier. But because there's no letter attached to the number, it makes the English comprehension of it all a nuisance.
  • dullard - Thursday, September 30, 2021 - link

    Ian. You are overthinking it. "The First Intel 4 EUV Chip" is accurate, understandable, follows Intel's desired naming, and doesn't mislead.
  • shabby - Thursday, September 30, 2021 - link

    Bingo!
  • m53 - Thursday, September 30, 2021 - link

    @dullard: All the naming police are active on Intel thread. But no one blinks an eye when people call TSMC N7 as 7nm.
  • shabby - Thursday, September 30, 2021 - link

    Then complain to intel, maybe they'll rename it again. When you tell a lie over and over again people will actually think it is 4nm when it isn't, come on Ian...
  • m53 - Thursday, September 30, 2021 - link

    Explain how this is a lie. When AMD used TSMC N7 and called it 7nm, was that a lie too?
  • shabby - Thursday, September 30, 2021 - link

    On tsmc's site they call it 3nm, 5nm and 7nm,
    Apple calls their chip 5nm 🤷🏼‍♂️
  • m53 - Thursday, September 30, 2021 - link

    Exactly. But according to your definition TSMC, Apple, AMD all are lying since there are no actual 7nm feature in it. They are telling "a lie over and over again people will actually think it is 7nm when it isn't".

    https://www.eejournal.com/article/no-more-nanomete...
  • GeoffreyA - Friday, October 1, 2021 - link

    These companies, all their pants are on fire.
  • Oxford Guy - Saturday, October 2, 2021 - link

    The tu quoque fallacy is flying fastly and furiously.
  • ZoZo - Friday, October 1, 2021 - link

    TSMC calls it 7nm themselves:
    https://www.tsmc.com/english/dedicatedFoundry/tech...
    So if AMD lied, they just repeated TSMC's lie.
  • ZoZo - Friday, October 1, 2021 - link

    By removing the misleading claims of nanometers, Intel is actually leading an improvement in the marketing of that industry. It would be a shame if even reputable press such as AnandTech were to be less virtuous than Intel's marketing department. If there were a bullet to bit, I'd say it's the one of adding "process" behind "Intel 4" every time the latter seems awkward by itself.
  • DrMrLordX - Wednesday, October 6, 2021 - link

    Never mind that Intel changed the name of Intel 7nm to Intel 4 specifically to confuse people. All that bad press about Intel's 7nm process being delayed? Boom! Gone. Intel is using Intel 7 to fab Alder Lake after all, and it's coming out in 2021! Yay Intel!

    It's great that you were able to bring this information about Loihi 2 to the public, but you're doing people a disservice when you assiduously fail to point out that Intel 4 is actually Intel's old 7nm node, and that it's exceptional that ANYTHING is being fabbed on Intel 4 this year since the node is so badly-delayed that Intel can't/won't fab any part of Ponte Vecchio with it (having opted instead to use TSMC N5 for those particular chiplets).
  • DrMrLordX - Wednesday, October 6, 2021 - link

    As an addendum, I do respect that you are no longer calling Intel 4 "Intel 4nm" but still, some context would REALLY have helped here as to why it's such a big deal that Intel 4 is being used. You said you're surprised by the announcement that Loihi 2 is on a pre-production version of Intel 4, but you haven't laid out WHY you're surprised. And it's important to do that since Intel has done such a good job of obfuscating what Intel 4 really is, how delayed it is, or you know, all that stuff that Intel doesn't want people discussing anymore.
  • DougMcC - Thursday, September 30, 2021 - link

    Maybe you could adopt the standard of calling it the Intel I4 process and see if it sticks.
  • TheinsanegamerN - Friday, October 1, 2021 - link

    Then it should be "intel is using it's 'intel 4' process". This would indicate the process is called "intel 4", rather then labeling it "intel 4nm", which it is not and only spread misinformation about intel's capabilities.
  • name99 - Saturday, October 2, 2021 - link

    Just call it I4.
    Intel is using I4. It will switch to I3, then to A20.
  • m53 - Thursday, September 30, 2021 - link

    7nm was the old name. It is now renamed as Intel 4 (commonly known as 4nm) as a part of their IDM 2.0 since it has equivalent performance and density of a hypothetical TSMC N4 (commonly known as 4nm) node. It makes the comparison easy with TSMC.
  • NickConrad - Thursday, September 30, 2021 - link

    I understand, I'm just trying to explain where I was coming from at this point. The good doctor set me straight.
  • m53 - Thursday, September 30, 2021 - link

    Some people still thinks TSMC 7nm or or Intel 7nm actually has any 7nm feature. They don't. These are purely marketing names. TSMC's marketing names were more aggressive than Intel's. Intel didn't bother since they were not a foundry. But now as a part of IDM 2.0 Intel is becoming a foundry. So Intel needs to explain to their foundry customers that their 7nm is better than TSMC N5 but worse than TSMC N3. The easiest way to do that is to rename it to Intel 4 aka 4nm.
  • Spunjji - Friday, October 1, 2021 - link

    I don't believe anybody genuinely thinks that. Anybody with any interest in manufacturing nodes will have heard that the names don't match the feature sizes some time back in the 2Xnm era, to the point of nausea now.

    Whether or not Intel 4 is better than N5 and worse than N3 remains to be seen - it's currently a marketing target. We don't even know how Intel 7 compares with TSMC N7 yet, save to say that if it has a similar density to previous Intel "10nm" nodes then it's probably not as dense.
  • shady28 - Thursday, September 30, 2021 - link

    For anyone that wants to see just how meaningless node by nm is, just go look up known densitys for say the 10nm 'node'.
    It ranges from 51MT/mm for Samsung 10nm to over 100MT/mm for Intel 10nm. But even these MT/mm numbers are sketchy - it depends on what the function of the circuit is exactly how many you can fit in a mm. Cache RAM for example is a lot more dense than a logic circuit.
    Hence nm node definition is meaningless. It's become a marketing term, nothing more and nothing less.
  • Oxford Guy - Saturday, October 2, 2021 - link

    Not meaningless but vague.
  • timecop1818 - Thursday, September 30, 2021 - link

    Just like FakeSMC 5 is like a 10nm node in reality. Are you going to bring up these bullshit numbers each time a bullshit node name is mentioned now?
  • yetanotherhuman - Friday, October 1, 2021 - link

    No, it's correct. Intel 4 is the name of the process. No "nm" in the name any longer.
  • DougMcC - Thursday, September 30, 2021 - link

    Wow, pretty amazing stuff. I was working on early neuron modeling software 30 years ago, and remember thinking it was pretty cool that we could model a handful of neurons at near real time.
  • Evan Shaw - Thursday, September 30, 2021 - link

    Is this the first real product to use asynchronous clocking? I know about a few projects that were done in the '80s that sounded really cool, but not much came of them.
  • yeeeeman - Thursday, September 30, 2021 - link

    Dr Ian, please correct the title. It is Intel 4, not intel 4nm
  • m53 - Thursday, September 30, 2021 - link

    I don't like people calling TSMC N7 as 7nm since there is no actual 7nm feature in it. But people always do that without anyone policing that name. That's why it is fair play to call Intel 4 as 4nm.
  • DrMrLordX - Wednesday, October 6, 2021 - link

    Everyone that I know of that posts on the forums refers to TSMC nodes by their actual names: N7, N7+, N7P, N5, N5P, etc.

    Intel was the last major foundry player to use the letters "nm" in their node names, until they changed node names to confuse people. Good job helping them in their goal by the way!
  • catavalon21 - Thursday, September 30, 2021 - link

    Ian wrote an entire article on how Intel is renaming nodes; it shouldn't be a surprise that an article with a title contrary to that information by the same author getting questioned.
  • m53 - Thursday, September 30, 2021 - link

    It is not contrary to that article. If he called it 7nm that would have been contrary.

    Calling it Intel 4 is great. But calling it 4nm isn’t contrary or incorrect. This is just what the industry used to hearing. It also helps laymen’s to understand this this node is slightly better performing than TSMC N5 (commonly known as 5nm).
  • shabby - Thursday, September 30, 2021 - link

    Whose payroll are you on?
  • m53 - Thursday, September 30, 2021 - link

    People only do personal attacks when they know they are wrong but don’t have enough courage to accept it.
  • shabby - Thursday, September 30, 2021 - link

    Please read intel's own documents https://images.anandtech.com/doci/16823/Intel%20Ac... there's no nm in their new naming scheme.
  • m53 - Thursday, September 30, 2021 - link

    Your concern was "When you tell a lie over and over again people will actually think it is 4nm when it isn't". Yet you were fine letting TSMC lie when calling their N5 process as 5nm. There isn't any 5nm feature in TSMC N5.

    Now regarding Intel's official naming, this is NOT an Intel official site. It's Anandtech. They are not required to follow intel's marketing guidelines. He just has to do a fair review of the process and make it easy to understand for the reader. In his previous article, Ian already explained how Intel 7 is similar to TSMC N7 and how Intel 3 is expected to be similar to future TSMC N3. Hence calling Intel 4 process 4nm makes it easier for the general reader to understand where the performance stands.
  • shabby - Friday, October 1, 2021 - link

    Good news... Ian changed it so no need so use reverse psychology on me.
  • catavalon21 - Thursday, September 30, 2021 - link

    My point was his prior entire article is how the naming convention is leaving "dimensions" behind - nanometers, angstroms, whatever. Just a number. And yes, the effort put into pointing all that out could have been spend doing far better things.
  • Spunjji - Friday, October 1, 2021 - link

    "It also helps laymen’s to understand this this node is slightly better performing than TSMC N5 (commonly known as 5nm)."

    We don't yet know this, though. All we know is that's where Intel expect it to land. They told us 10nm would be around 50% more dense than TSMC N7, and it isn't.
  • Lakados - Monday, October 4, 2021 - link

    We already know that Intel's 7nm process is roughly equivalent to the TSMC 5 as Intel has been using TSMC's 5 to cover their own 7nm shortfall with the Aurora supercomputer. From internal reports, the Intel 7nm (now Intel 4) performed nearly identically with the TSMC produced chips to the degree where the only discernable difference was the SKU numbers on the card that identified them as different.
  • eastcoast_pete - Thursday, September 30, 2021 - link

    Regardless of what size the structures really are, I took two things away from this article: 1. These neuromorphic processors are cool tech, something to watch, and 2. Intel is seriously gearing up for EUV in their manufacturing. Point 2 is good news for all of us, because I fully expect AMD moving its next generation to TSMC's 5 nm or better to keep up. Interestingly, that leaves NVIDIA bringing up the rear in what node they are on. Samsung 8 is far from either TSMC 7 or Intel 7.
  • Spunjji - Friday, October 1, 2021 - link

    Last I'd heard, Nvidia were considering moving back to TSMC. Still, if they go with Samsung 5nm for their next products, they'll get something close to TSMC N7 in electrical performance but with higher density (and less competition for wafers). That might work reasonably well for them given the generally high energy efficiency of their architectures.
  • wrkingclass_hero - Friday, October 1, 2021 - link

    Thank you Ian for clearing up the title.
  • Spunjji - Friday, October 1, 2021 - link

    I found this very interesting, both in terms of being a novel area of computing I'm not familiar with and to see what Intel have chosen as a "pipecleaner" for their next gen process. It's certainly a more interesting project than printing out little dual-core CPU dies!

    Cheers for the good content, Ian.
  • Diogene7 - Friday, October 1, 2021 - link

    @Ian: I was wondering if there would be any chance to get an article that would compare the new Intel Loihi 2 chip versus IBM Synpase Truenorth chip, because I think both are roughly 1 million neurons chip ?

    What are the architectural differences and what each of them can / can’t do ?

    I would be very curious to know…
  • Juneli - Sunday, October 3, 2021 - link

    I develop with the dataset of the neuromorphic chip Akida from Brainchip. Real Spiking Neural Network and simple programming. Even one-shot learning at the edge without cloud works! An incredible asset for developers!
  • Wereweeb - Friday, October 1, 2021 - link

    Out of all the things to be debated, the internet chose naming conventions. Bo-ring.

    I am both excited and terrified by this. We could be coming closer to true AI.
  • GeoffreyA - Saturday, October 2, 2021 - link

    In typical human fashion, we debated the frame, instead of looking at the picture. Alas.
  • Oxford Guy - Saturday, October 2, 2021 - link

    People obsess about it because of stock investments. They take marketing magic very seriously.
  • GeoffreyA - Saturday, October 2, 2021 - link

    Thanks for a great article, Ian, as always. To me, it looks as if this technology is one of the steps on the road to consciousness.
  • Oxford Guy - Sunday, October 3, 2021 - link

    The conscious computer will take one look at human behavior and suffer cascade failure like Lal.
  • Oxford Guy - Saturday, October 2, 2021 - link

    Ricoh can use it to build those copiers and other office equipment that use AI to force workers to smile genuinely in order to get the equipment to function.

    A joyous future awaits.
  • GeoffreyA - Sunday, October 3, 2021 - link

    Smiling CEO in Silicon Valley: "A better future. For all."

Log in

Don't have an account? Sign up now