Intel Optimistic About Its Next-Gen 7nm Process Technology (anandtech.com) 114
From a report: Originally planned to enter mass production in the second half of 2016, Intel's 10 nm process technology is still barely used by the company today. Currently the process is used to produce just a handful of CPUs, ahead of an expected ramp to high-volume manufacturing (HVM) only later in 2019. Without a doubt, Intel suffered delays on its 10 nm process by several years, significantly impacting the company's product lineup and its business. Now, as it turns out, Intel's 10 nm may be a short-living node as the company's 7 nm tech is on-track for introduction in accordance with its original schedule.
For a number of times Intel said that it set too aggressive scaling/transistor density targets for its 10 nm fabrication process, which is why its development ran into problems. Intel's 10 nm manufacturing tech relies exclusively on deep ultraviolet lithography (DUVL) with lasers operating on a 193 nm wavelength. To enable the fine feature sizes that Intel set out to achieve on 10 nm, the process had to make heavy usage of mutli-patterning. According to Intel, a problem of the process was precisely its heavy usage of multipatterning (quad-patterning to be more exact).
For a number of times Intel said that it set too aggressive scaling/transistor density targets for its 10 nm fabrication process, which is why its development ran into problems. Intel's 10 nm manufacturing tech relies exclusively on deep ultraviolet lithography (DUVL) with lasers operating on a 193 nm wavelength. To enable the fine feature sizes that Intel set out to achieve on 10 nm, the process had to make heavy usage of mutli-patterning. According to Intel, a problem of the process was precisely its heavy usage of multipatterning (quad-patterning to be more exact).
This is where Intel re-labels. (Score:5, Interesting)
And before you go there, Intel was the first company to lie about node size.
Re: (Score:2)
Yeah, I just figured that they were going to put a header line at the top of all of their tapeout files:
Re: (Score:2, Troll)
In the era where there was a fairly broad consensus about how to advertise a process node size, there was only one transistor.
But then the power envelope started to matter, and signalling distance increasingly started to matter, and then there wasn't just one transistor any more, but various transistors tuned for different roles, and some transistors are contributing to peak clock frequency, while others are contributing to total in
Re: (Score:2)
No idea what happened there. Truly spastic.
Re: (Score:2)
Re: (Score:2)
Intel will now re-label their 10nm as 7nm.
Yes they will. And they will just be replacing some of the multipatterning steps of current 10nm node. If they claim the EUV team is working independently then they are lying. And if they end up needing EUV to bail out their current 10nm node then they are in for an even bigger disaster than they already have, because EUV is still months away from anything resembling high volume.
Re: (Score:2)
For those that didn't read the article, Intel's 7 nm technology was developed independently by a separate internal team than the 10 nm technology. The 7 nm tech avoids multi-patterning, which is apparently the part that gave rise to the yield problems with the 10 nm.
Re: (Score:2)
Basic business question before that: to what extent will 7nm be a cost reduction on whatever Intel had before it? Those $10B fabs would have to be factored in whatever cost calculations they make on the new nodes, which I'll wager would actually cost more than current nodes, and would only be profitable in an allocation market
Good Time to Stop Hardware Obsolencene (Score:5, Insightful)
Re: (Score:1)
It is 7 years old for fucks sake. You obviously don't need more cores or you would have upgraded. No one is going to support your old shit forever.
Re: (Score:1)
That motherboard of yours is 12 to 18 months from failure. Electrolytics don't last forever. And while the caps used by reasonable manufacturers aren't the worst on Earth, they all have lifespans measured in a small multiple of ten thousand hours.
Re: (Score:1)
I had a P166MMX from 1997 with 32MB RAM that runs just fine. Finally got rid of it this year. The wife was using it for Word/Excel and Email and was upset I decided to retire it. It was on 4 or 5 days a week since it was new. In that time a few HDDs have come and gone (ending with a CF card running Win98SE), two CRT monitors (also retired) and a LCD. The Ethernet card had to be replaced, but the analogue modem worked until the end (tested it on a fax just to see). Several mice, but the original keyboard is
Re: (Score:2)
Re: (Score:2)
I need more cores, but I am not throwing this motherboard, chip, and memory out, any time soon.
You must be new to using Intel CPUs. They use a new CPU socket every 18 months, no upgrades for you.
Re: (Score:3)
That Gamers Nexus 2017 i7-2600K benchmark has been updated for 2018: i7-2600K benchmark. [youtu.be]
i.e.
They added Threadripper 2990WX metrics both in Creator Mode and Game Mode.
Sad that modern CPUs only give single digit performance increases. :-/
Glad you mentioned Civ 6 -- a perfect example that shows some games are still dominated by single core performance. (Quad core has been a standard for *years* but some game devs are still not using it efficiently. Granted not everything is parallelizable but console g
Re: (Score:2)
AMD's gaming mode disables half the cores because half of them have a crippled memory bus. The four chip package only has two memory controllers. Two of the chips need to access memory via the other two.
Re: (Score:2)
I need more cores, but I am not throwing this motherboard, chip, and memory out, any time soon.
Of course having a bunch of old hardware that you don't really need lying around is so useful... if you're just looking at the environmental angle and not the economic one just sell it and buy what you actually need as somebody else will be actually using it. Personally I notice I'm a bit too optimistic and lazy and indifferent, I buy something and then realize it's not really as useful as I thought or I never get around to using it or I used it before but don't really need it much anymore but the second ha
Re: (Score:1)
I mean, I'm sorry man but that simply isn't true. Sandy Bridge is great, but particularly for media workloads, my Coffee Lake absolutely wipes the floor with the _Ivy Bridge_ it replaced, to say nothing of things like single-core performance for tasks like emulation. You're probably right that for ordinary, everyday MS Office and Internet stuff it's not that big a deal, but to say 7 years of development results in a 30%
Re: Good Time to Stop Hardware Obsolencene (Score:1)
Your ancedote doesn't match the numbers tho. The benchmarks say only a 30% increase. Go check them.
Re: (Score:2)
Perhaps the same benchmarks only show minor gains because they don't use the new instructions the more modern CPU's support.
Re: (Score:2)
It seems contrived and misleading of you to limit your comparison to current quad-core CPUs, as Intel hardly makes those anymore. If instead comparing the launch cost of the 2600K [cpubenchmark.net] (~300 USD), then a current match would be the 8700 [cpubenchmark.net].
For that you get approx. +80% multi-threaded performance and +35% single-threaded performance, for -30% power consumption.
Re: (Score:2)
I partly disagree.
The benchmark [youtube.com] by Gamers Nexus shows that there is a ~+35% improvement for Civ 6 turn time for 8700K (similar performance to 8700, except overclockable). This is not nil.
Also, GP did not specify his workload, so I gave him generic benchmarking results which seemed to contradict him.
So I still consider GP to have made a misleading comment by making a contrived comparison.
With that said, I agree with you that current CPU performance may not motivate people to switch. That is however a judgeme
Re: (Score:2)
Turn time was 15.7 for 2600k and 12.5 for 8700k. So, I stand corrected that it's not nil. But it's closer to ~+25%.
Thanks for the correction. I must've looked at the overclocked number there.
Um, if you don't know his workload, using a generic benchmark doesn't contradict him. It only proves that if he *has* a generic workload that his statement is contradictory. That's why I pointed out the Gamers Nexus benchmark because it's possibly much closer to his workload.
I was careful to use the word seems to show that this was my impression based on his unspecific "30% slower" comment. Other readers less versed in the nuances of measuring CPU performance could get the impression that there actually only is a 30% performance difference between old and new CPUs, which I consider to be a misleading measure. This is why I challenged him with more specific numbers, which gives another picture.
YMMV. If you want to argue that the GP was unclear because he didn't specify his workload, that's a potentially valid complaint. The problem is, if he states on the order of a 30% improvement, that's very much in line with the likes of Gamers Nexus results. I'd tend to argue that statement was precisely in reference to those who have been paying attention to results in games and seeing similar numbers. It's not contrived to suggest that most gamers, in the same boat, would understand the comparison being implicitly made.
I also ass
Re: (Score:2)
My 7-year-old intel 2600k is only 30% slower than current quad-core offerings.
Aside from that number being 40% what else can your 7 year old 2600K do?
Does it have 24x PCI3.0 Express lanes? Or are you runnin PCIe 2.0 at half the speed and 1/3rd of the available lanes, barely enough to power a graphics card and a modern SSD? What else do you think sucks about modern CPUs? Double the memory bandwidth? Native NVMe interfaces to SSDs for orders of magnitude faster bandwidth compared to SATA? 2nd generation processors, you must also be a fan of USB 2.0 and it's 1/20th the bandwidth of USB
Re: (Score:2)
Even the electrolytic capacitors have longer lifespans now. But that's not your only option.
Replacing the power supply is plenty cheap. But for the same reason above, you may not have to for a long time. It's the CPU fan that you'll probably have to change first.
They better be optomistic. (Score:3)
The 10nm delays were hailed as "the end of intel!" and "AMD will destroy them" and other kinds of ridiculous things, over a year ago.
A year later and I'll say, these delays are very significant now, it's very likely that AMD will genuinely have a processor out that has superior process technology in it. May not be faster due to design but the gap closing faster.
Also, Intel 7nm chips will be superior to other "7nm" chips out there, since companies kind of just make up the numbers now. Intel 10nm = TMSC 7nm basically. Quite frustrating.
Regardless Intel needs this, 14nm is seriously long in the tooth for Intel and there are very obvious limitations to what can be done for efficiency, performance, they're very much running out improvements.
Re: (Score:2)
Big business doesn't care about faster single core (Score:3, Interesting)
AMD already has a *much* better processor out there. Because businesses calculate in dollars per performance. They like as many cores as possible with the fastest interconnects.
In all of those aspects, Intel currently is a joke, compared to AMD.
Which is exactly, why Epyc processors are currently making big cuts into Intel's Xeon revenues.
Intel is currently in panic mode.
It's really pathetic, how Intel fanboys always cling to the one thing they have left: Meltdown-included single core performance.
By acting a
Re: (Score:2)
Consoles have been multi-core for a long time
13 years ago, XBox 360: 3 core IBM CPU
12 years ago, PS3: 8 core IBM CPU, 6 used for running the game.
6 years ago, Wii U, 3 core IBM CPU
5 years ago, PS4: "8" core AMD CPU
5 years ago, XBox One: "8" core AMD CPU
Re: (Score:2)
End user desktop benchmarks, do not correlate, in any way, with your post for average use case. Single core is still king in a plethora of tasks, and likely will be for a long time.
Server performance, on seriously threaded applications, per $ per core, $ per watt, etc, AMD is doing amazing, I'm very happy for them and it's awesome.
This changes nothing for home users. Intel still hammers it home at single core, not just games. They have better IPC and more frequency.
I am not intel, I do not own intel sh
Re: (Score:2)
Oh and this specific comment of yours, is wildly incorrect
">And no, Intel's "7nm" will not be superior to other "7nm" chips, because it WILL be Intel's 10nm process! ^^
(Intel came up with it, btw, according to the current top comment.)"
Intels 10nm = similar to TMSC's 7
This article is specifically about INTEL doing their own refresh to 7nm as in "intel 7nm" not "generic 7nm" - this will likely be similar to what TMSC claim to be 5nm.
If you have any understanding of how this stuff works or follow tech news
Re: (Score:2)
This article is specifically about INTEL doing their own refresh to 7nm as in "intel 7nm" not "generic 7nm" - this will likely be similar to what TMSC claim to be 5nm.
I actually doubt that. I think it will be Intel just renaming their 10nm as 7nm and adding some EUV mask steps to justify it. So the days of endlessly re-explaining that "Intel 10nm is similar to TSMC 7nm" will be over, and that's all that will change. Intel is not going to magically leapfrog TSMC and retake the process lead.
Re: (Score:2)
They have been working on 7nm in parallel though, so it's not out of the question they could pull something out of the bag.
Yes it is. Intel is using exactly the same ASML equipment as TSMC and Samsung. It is a very safe bet that they are using it at exactly the same resolution.
BTW, as of today, EUV is not anywhere close to ready for high volume production, only samples. One huge issue is pellicles, those things that cover masks to keep them clean. The problem is, all matter is opaque to EUV so the pellicle needs to be made ridiculously thin to pass enough light, and has to not burn up given the megawatt or so they pump into the
Re: (Score:2)
"It is a very safe bet that they are using it at exactly the same resolution."
Even if the EUV tool has the same resolution (theoretical upper limit is 41 nm pitch for parallel lines for EUV exposure at NA 0.33 and 13.5 nm wavelength; 71 nm pitch or so for 193 nm at NA 1.35), the resulting feature size in the chip is hugely dependent on the (etching and deposition) process that follows. Intel/TSMC/Samsung have specialized in very different processes. It's amazing how process techniques can result in features
Re: (Score:2)
"theoretical upper limit is *21* nm pitch for parallel lines for EUV exposure at NA 0.33 and 13.5 nm wavelength; "
Sorry, forgot a factor 2. There, FTFM.
Re: (Score:2)
According to ASML's reports last summer, the EUV source outputs 250 W. Are you talking about electrical power?
Right. I confess to a bit of conflation there, to bring home the full scary picture of an EUV stepper. Hardly your granddaddy's gestetner. Semiwiki says 300 watts [semiwiki.com] of EUV is desirable for high volume. According to Wikipedia the input power rounds up to a megawatt. [wikipedia.org]That doesn't directly heat the pellicle of course, but it does require a small creek's worth of cooling. The pellicle is relatively late in the optical path so only a fraction of the 300 watts goes through it (twice) but it's still a huge problem fo
Re: (Score:2)
Wikipedia says ASML's current offering, the NXE:3400B, has 13nm resolution. My (shaky) understanding is, that means barely capable of single patterning at 7 nm. It looks to me like EUV double patterning will be a thing already at 7nm, and certainly at 5nm. But at least it's better than quad patterning that reportedly defeated Intel.
I think everybody's in the same boat, basically, and nobody has a magic bullet. Intel thought they were smart enough to make quad patterning work where others respectfully shied
Re:They better be optomistic. (Score:4, Interesting)
The 10nm delays were hailed as "the end of intel!" and "AMD will destroy them" and other kinds of ridiculous things, over a year ago.
It's funny you say that because that's happening right now. You can claim an Intel core is faster but when you factor in Specter and Meltdown mitigations then it's slower, especially because server hosts disable SMT entirely for Intel CPUs. Gamers may not be interested in tight security but that's exactly what the server world wants and that's where the real money is.
Right now, Intel is hemorrhaging money to try to keep up their PR/anti-competitive game in hopes that they will be able to come out with a core without such flaws before the damage is too much to come back from. This bullshitting is just part of their PR game.
Re: (Score:2)
So you're basically entirely incorrect
Even with spectre, Intel single core is decimating amd.
It's a shame, I like competition, but you're simply wrong.
Re: (Score:2, Interesting)
That's not the case. In fact, that's horseshit starting from a flawed inception, that all IPC measurements are taken equally. They aren't. Intel chips dominate AMD in terms of forcing developers to optimize for them exclusively.
Such is the case with several benchmarking tools that have been exposed and "known" to do this for some time.
They don't really demonstrate a real-world measurement, they demonstrate a business dev budget in the billions.
In the real world, Intel has something around a 7% lead in IP
Re: (Score:2)
Not only is this a known fact, Intel lost a god damned lawsuit on this very issue.
Re: (Score:2)
In the real world, Intel has something around a 7% lead in IPC for the average load which includes some arguably Intel-optimized testing and some (most) agnostic.
Right, in return for a 60-100% advantage in price per core. Not hard to see why many gamers are deciding to spend those savings on the GPU.
Re: (Score:2)
Also, Intel 7nm chips will be superior to other "7nm" chips out there, since companies kind of just make up the numbers now. Intel 10nm = TMSC 7nm basically.
Any other Intel P.R. you want to spew?
Intel doesnt really have what Intel would call 10nm yet. There is no mass production, only test runs continually proving that mass production aint possible yet. If mass production aint possible, its just a showpiece. There is no way to know what transistor densities Intels 10nm would have because they dont have 10nm.
TSMC on the other hand is ramping up 7nm mass production as we speak. Intel now has a target and if that target is what they would call 10nm then they
Re: (Score:2)
It's an outright fact that intel measures their foundry processes FAR closer to real figures than the others. The numbers have become almost entirely meaningless for several years now.
10/12nm from the competition is basically the same or slightly worse than intels 14nm
TMSCs 7nm IS very much similar to Intels 10nm. This isn't Intel PR, this is as measured, documented, reported by HUNDREDS of technical sites which follow this stuff.
Intel may well be well behind on 10nm, their 10nm may even be designed poor
Re: (Score:1)
It's an outright fact that intel measures their foundry processes FAR closer to real figures than the others
What real figures? There is no 14nm dimension in Intel's 14nm node [wikichip.org] either. Intel's node name may be more conservative than TSMC's, but nothing makes it "real". You are just blowing smoke out your ass.
Re: (Score:2)
You are quoting
> "It's an outright fact that intel measures their foundry processes FAR ***__closer__*** to real figures than the others"
With
> " Intel's node name may be **__**more conservative**__** than TSMC's"
(my emphasis)
So thanks for proving my point I guess?
Are you very confused? You seem to be.
Re: (Score:1)
I am not confused about whether you are an ass or not. You claimed that Intel's process name has something to do with real numbers. It does not. My point.
Now go ahead and wank on about whether Intel's node name is conservative or not if you must. I did not debate that.
Re: (Score:2)
Mate, go away. You're embarrassing yourself. Just stop. Go abuse someone else. /ignored.
Re: (Score:1)
Yet another content free post from you. Got any real dimensions to quote, or is there nothing about you except foul mouthed rhetoric?
Re: (Score:2)
Re: (Score:2)
What part of "real" do you not understand? If you think that Intel's node name has something to do with any real dimension then please educate us about which dimension that is.
Re: (Score:2)
What part of "closer" do you not understand?
Re: (Score:2)
Also, Intel 7nm chips will be superior to other "7nm" chips out there, since companies kind of just make up the numbers now. Intel 10nm = TMSC 7nm basically.
I'll take issue with that. The other two players are basically going to introduce EUV at the same key dimensions as their current deep EUV node, to replace some costly and error prone multipatterning steps. I would not be surprised to see multiple iterations here, until nearly all deep UV is replaced by EUV. That is all going to happen before 5nm has any chance at all of becoming reality.
For Intel to do otherwise would be the height of idiocy, practically ensuring that 7nm turns into a repeat of the 10nm fi
Mind-bending (Score:3)
Comment removed (Score:4, Informative)
Re: (Score:3)
Ok, we won't quote you, random internet guy.
Re: (Score:2)
Re: (Score:2)
Given that the diameter of a silicon atom is around 0.2nm, that means they are now building transistors out of something like 30-35 atoms across. How far down can this go before it all disappears in some kind of quantum uncertainty blob?
As I understand it there are few enough atoms that allegedly Micron has named them. Not kidding.
Re: (Score:2)
Given that the diameter of a silicon atom is around 0.2nm, that means they are now building transistors out of something like 30-35 atoms across. How far down can this go before it all disappears in some kind of quantum uncertainty blob?
New Metal-Air Transistor Replaces Semiconductors. A novel field emission transistor that uses air gaps could breathe life into Moore’s Law
The ACT device eliminates the need for semiconductors. Instead, it uses two in-plane symmetric metal electrodes (source and drain) separated by an air gap of less than 35 nanometers, and a bottom metal gate to tune the field emission. The nanoscale air gap is less than the mean-free path of electrons in air, hence electrons can travel through air under room temperat
Re: (Score:2)
iPhones already have problems with a little bit of helium in the air with their mems clocks.
Even hours of exposure to levels that don't make you sound like a chipmunk can kill the phone to the point it takes days before the He makes its way back out the "sealed" chip.
Re: (Score:2)
Re: (Score:2)
Given that the diameter of a silicon atom is around 0.2nm, that means they are now building transistors out of something like 30-35 atoms across. How far down can this go before it all disappears in some kind of quantum uncertainty blob?
Functionally, if you're looking just to produce one transistor and not billions of them then around 3 nm:
In 2006, a team of Korean researchers from the Korea Advanced Institute of Science and Technology (KAIST) and the National Nano Fab Center co-developed a 3 nm transistor, the world's smallest nanoelectronic device based on conventional finFET technology.
Of course if you choose exotic designs even a single atom can function as a transistor or possibly even electrons and whatnot. But compared to the current gen 7nm chips that TSMC got in the iPhone etc. there's two more generations, 5nm and 3nm then anything resembling current chips is done. If they can even make that work in volume, though they've managed to continuously pull new rabbits out of the hat.
Re: (Score:2)
Given that the diameter of a silicon atom is around 0.2nm, that means they are now building transistors out of something like 30-35 atoms across.
Thats not how it works and it hasnt been that way for many generations.
The pitch of Intels latest "14nm" process is 42nm and the gate lengths are 20nm. These are the key values that indicate the transistors performance. The latest have a width of 8nm, which effects density but not so much performance.
Contrast that with Intels latest "10nm" process, which has a superior pitch of 34nm, but an unknown/unreported gate lengths. These test chips are slower than Intels latest 14nm so it is probably safe to as
Re: (Score:2)
Given that the diameter of a silicon atom is around 0.2nm, that means they are now building transistors out of something like 30-35 atoms across
Not even close. First, the silicon cubic lattice side is 0.543 nm, the atomic spacing is roughly 1/3 to 1/2 of that. When counting atoms, that's what you use, not atomic radius (which you seem to have misquoted as diameter.) Second, the gate pitch for TSMC's 7nm is 54 nm [wikichip.org] So each transistor covers an area something like 100-150 atoms on a side on the very crude assumption that the transistor is square, or about 10K to 40K atoms.
Somebody who actually does this for a living, please check my numbers.
shady tactics (Score:1)
Re: (Score:2)
Because they were building both 10 and 7 at the same time and they use completely different technology to do each one?
Makes sense.
10nm used multi-patterning and 7nm is using narrower wavelengths.
It's understandable that delays in one may not affect the other.
CEO (Score:2)
Who is CEO of Intel these days? Ballmer?
It seems that they are really dropping the ball!!
Intel behind on their chip fabs?
Major security flaws in their chips?
Is this even possible for the once dominating chip company?
ps. I thought Ballmer was at Apple these days.. but it seems he's CEO of both spending more time at Intel?
We all know why Intel is making this announcement (Score:1)
The leaked AMD Ryzen specs spooked them, simple as that.
ASML (Score:1)
As someone who works on EUV at ASML, I'm getting a kick...
This is sort of like Bhutan Airways claiming that they set the world record for the longest flight, when it's the airplane that was made by Boeing...
No one will have true 7nm "node" without EUV, and no one has machines that do EUV lithography besides ASML. Intel knows this or they wouldn't have given ASML a ton of money.