Slashdot Mirror
Intel Has Killed off the 10nm Process, Report Says (semiaccurate.com)
Charlie Demerjian, reporting for SemiAccurate: SemiAccurate has learned that Intel just pulled the plug on their struggling 10nm process. Before you jump to conclusions, we think this is both the right thing to do and a good thing for the company. For several years now SemiAccurate has been saying the the 10nm process as proposed by Intel would never be financially viable. Now we are hearing from trusted moles that the process is indeed dead and that is a good thing for Intel, if they had continued along their current path the disaster would have been untenable. Our moles are saying the deed has finally been done.
This isn't to say the road to this point has been easy or straightforward, and the road ahead is even less solid. Intel has continually moved the public bar on 10nm back, incrementally, while singing a different song internally. In their Q1/2018 earnings call they moved the timetables and spun it in a curious way but were telling partners a different story. UPDATE: Intel tweeted on Monday morning: "Media reports published today that Intel is ending work on the 10nm process are untrue. We are making good progress on 10nm. Yields are improving consistent with the timeline we shared during our last earnings report."
This isn't to say the road to this point has been easy or straightforward, and the road ahead is even less solid. Intel has continually moved the public bar on 10nm back, incrementally, while singing a different song internally. In their Q1/2018 earnings call they moved the timetables and spun it in a curious way but were telling partners a different story. UPDATE: Intel tweeted on Monday morning: "Media reports published today that Intel is ending work on the 10nm process are untrue. We are making good progress on 10nm. Yields are improving consistent with the timeline we shared during our last earnings report."
Prepared how? You say it like it's gonna change anything.
Interesting move. Time will tell if this move is more in the interest of shareholders and will hinder technological advancements...
"In a time of universal deceit, telling the truth is a revolutionary act." - George Orwell
There are plenty of gains to be had by optimizing all the sloppy and inefficient code that has been written and re-used through the years which was accepted thanks to constant hardware improvements.
I'm not an expert in semiconductor fabrication. That said, if the issue with 10nm was due to incomplete fabrication tooling and/or management, that's one thing. But I fail to see how Intel will fare better with 7nm or 5nm if the problem was purely of physics. If it's the later, we might have effectively hit the wall in terms of being financially feasible.
Life is not for the lazy.
There are plenty of gains to be had by optimizing all the sloppy and inefficient code that has been written and re-used through the years which was accepted thanks to constant hardware improvements.
Absolutely true. However, hardware improvements are a one time cost that makes all your software faster. Improving software speed is a painstaking process of going through and re-writing code.
In other words, hardware improvements are practically free. Software optimization is incredibly expensive. My expectation is that HW development will move towards more and more specialized HW improvements.
If what the report says is correct
This isn’t to say the road to this point has been easy or straightforward, and the road ahead is even less solid. Intel has continually moved the public bar on 10nm back, incrementally, while singing a different song internally. In their Q1/2018 earnings call they moved the timetables and spun it in a curious way but were telling partners a different story.
Nothing however tops the masterful “Hyperscaling” stunt where Intel brought in press and analysts to a ‘manufacturing day’ in early 2017 to explain how the crippling slide of 10nm was not actually a slide, it was a good thing and not a delay at all. SemiAccurate laughed and stopped just short of calling Intel liars.
The company redefined terms well past the breaking point to show that scaling was ‘on track’ even if node cadence was ‘intentionally’ longer. As you can see from the above graph, all was good publicly, internally SemiAccurate was hearing a very different story. (Note: Intel was on track to miss that graph by 1+ year and sliding before 10nm was killed.)
Be interesting to see how this plays out.
I find it hard to take any article seriously that uses the word "moles" multiple times that isn't discussing a skin issue.
Intel hasn't killed off anything. This idiot Charlie Demerjian quotes his own site to "prove"...what?
fact is Intel is delaying 10nm (a marketing term more than an actual chip feature size, doesn't mean half pitch any more) to 2019
Seems to be doing just fine with their venture into 7nm land.
There are chips in development that are 1000x more efficient (performance for the power) compared to current GPUs for neural network training and playback in development, and the development chips are designed on existing processes. See the paper on RapidNN for an example.
The secret to the speedup is to place memory and arithmetic units on the same chip. That technique has room for improvement still as memory and logic processes are a little different and combining them on the same chip results in inefficiencies.
We are approaching the limit for CMOS-based electronics, for many years there has been a lt of research for an alternative, although none are here yet. But when one of those alternatives finally arrives, it will be a second revolution of computing, like being back in the 90s. One promising technology is vacuum channel transistors, for example.
...or research new algorithms. Nah, can't have that!
Ezekiel 23:20
It's very likely that we won't see 7nm or 5nm for another few years
Actually, you'll probably see 7nm next year.
Ezekiel 23:20
If for the same die that's the part they have problem with.
If you just mean use more of everything sure that give more performance but not with the cost and energy cuts.
End of the line for ever more powerful digital computing is coming fast. Better be prepared.
While Moore's law is definitely in big trouble, the spirit of what it hoped to offer may not be. The cost of chips continues to come down. The power consumption continues to come down. While the transistor density may no longer be increasing as fast, they continue to increase CPU core counts on larger dies.
CPU chips for most real world applications haven't been the bottleneck for a while. Solid state non volatile storage continues to advance and this is the primary remedy for most application slowness.
Machine learning technology may be the future of what people want for ultra powerful low cost computers. Architectural advances can offer vastly more bang for the buck than raw transistor count. Not only to graphic chip architectures offer this but extensions to Intel machine instructions like AVX have a more dramatic impact than the tick/tock of Moore's Law
IMHO the death of Moore's Law != the death of more powerful computing.
Greed is the root of all evil.
Hey, it's good for job security. I shouldn't say that because I often rant about the how stupid web (non) standards turned simple inhouse CRUD dev into a labor-intensive Dagwood Sandwich mess and fashion contest.
Table-ized A.I.
Ummm.... You people do know that Semiaccurate is a kind of notorious for being run by an insanely zealous AMD fanboy? We're talking about the same person who made some absolutely crazy misreading of a number of documents to arrive at a conclusion that Nvidia's Fermi architecture had sub 5% yields as devices based on that architecture were shipping in good numbers.
A reliable source this is not, particularly when it comes to AMD's or it's direct competitors, so wait for some more trustworthy sources before making your mind up on this subject.
"Why should I want to make anything up? Life's bad enough as it is without wanting to invent any more of it."
New algorithms? New algorithms might just introduce new unintended consequences because coders aren't perfect. How about cleaning up and fixing old codes that introduce opportunities for hackers.
In a time of universal deceit, telling the truth is a revolutionary act. George Orwell
TSMC have been in 7nm risk production mode a long time (for the semiconductor industry), or so they claim.
Except they won't be cards, they just would be IC embedded on the main board, or even within the same chip.
I doubt we will see the return of the popularity of 3 foot tower PC. Loaded to the brim with expansion ports,
If something is so important that you feel the need to post it on the internet... It probably isn't that important.
To quote this August article About Intel's 10nm Process Lead:
The thesis of the article is that Intel was the first to attempt Self Aligning Quad Patterning (SAQP), in a "10mm" node that's more aggressive than the current or just getting going maybe Samsung and TMSC "7mm" nodes.
Per the author, Intel's "7mm" node will use EUV, which has its own host of problems, but is a fundamentally different technology, see the above link on pushing 193 nm immersion lithography, vs. using what are very close to being soft X-rays at 13.5 nm (per Wikipedia the X-ray range starts at 10nm). Of course, if Intel's 7nm node expected to get some use out of their 10nm SAQP technology, it still might be in trouble.
Parallelism has been on the drawing board for over a generation in prep of the end of Moore's Law, Within the past decade parallel cores have became popular. My Laptop Today with a good video card with nearly 2000 Cuda cores, give more parallel computing power then a MassPar mainframe With (1024 CPUs) I used back in my college days, learning parallel processing.
If something is so important that you feel the need to post it on the internet... It probably isn't that important.
Clearly this is a situation where there's no old code that could be exploited because it has never been put into practice.
Ezekiel 23:20
But I fail to see how Intel will fare better with 7nm or 5nm if the problem was purely of physics. If it's the later, we might have effectively hit the wall in terms of being financially feasible.
I don't think it's physics as much as financial. The demand is currently not there. A 5 year old laptop was faster than most current laptops/desktops/tablets/phones. Most people have stopped demanding high performance CPUs. Even gamers have shifted their focus to the video cards not the CPUs and a majority of games are now written for tablets/phones that have little more processing power than a 486. Because of the low processing power and low energy requirements of phones/tablets, the stuff that might have created demand like voice recognition has shifted to the cloud instead.
End of the line for ever more powerful digital computing is coming fast. Better be prepared.
There is plenty of room for improvement just very few niche applications to pay for it. The demand is currently not there. A 5 year old laptop was faster than most current laptops/desktops/tablets/phones. Most people have stopped demanding high performance CPUs. Even gamers have shifted their focus to the video cards not the CPUs and a majority of games are now written for tablets/phones that have little more processing power than a 486. Because of the low processing power and low energy requirements of phones/tablets, the stuff that might have created demand like voice recognition has shifted to the cloud instead.
That all depends on whether or not Moore's Law is truly dead, or only dead for this Intel "10nm" process, with the added wrinkle that non-recurring engineering (NRE) costs have been skyrocketing with these more extreme 193 nm immersion lithography nodes because multipatterning requires many more masks. In the latter case, customers who aren't going to order as many units won't use the newer nodes, unless they achieve some other need of their's, like lower power consumption.
Moore's Law has more been a "law" of economics, in that smaller nodes were for a very long time clearly cheaper to manufacture, since defects were random, especially the classic spot of dust types, and the more die you could fit on a wafer, the higher your yields. Here, Intel has conspicuously failed to get acceptable yields for years, at some point if they don't succeed they will have to pull the plug, either to a less aggressive node, or to the next one if they can get it to work and it's economic. Their "7nm" node is said to use EUV heavily, so the last question is also very much up in the air.
Be prepared for what? We reached a plateau of power vs need a long time ago. The average consumer the plateau topped out about 10 years ago. I am a power user and I'm having to make up excuses to upgrade my 6700K to something more powerful. It does everything I need it to do and with plenty of processing room left over.
I read at +2. If your post doesn't reach that level I will not see or respond to it.
Exactly. AMD appears to be betting the company on their 7nm process shipping next year. I've heard that AMD 's 7nm is roughly equivalent to Intel's 10nm. Anyway - the idea that there's no incentive anymore to shrink CPU dies anymore is dead wrong. Two things prevent CPU cores from processing faster and more efficiently - heat and the speed of light (or the speed of electrical signals in sillicon). Both are helped by using a smaller process. Almost all purchasers of CPUs benefit from this. On a laptop you can get better performance, better battery life or a smaller battery, or some combination of these. In a data center you can get more servers in a smaller space, fewer servers, lower cooling costs, etc. I don't believe for a second that Intel is going to stop trying to shrink their process - this is just a setback for them and if they don't act quickly, AMD will eat their lunch.
mac pro now delayed to 2020 (intel may it so AMD is not going to happen)
I'm struggling with what you're trying to say here. "may it so"? "AMD is not going to happen?" Huh?
I don't think Apples "tech refresh" is in any way dependent on Intel's process size. There hasn't been any correlation that I've seen between process size and "mac[book] pro" releases. Despite not yet hitting 10nm, Intel has still released new iterations of their i-series chips pretty consistently.
Intel has done some underhanded things in the past to push AMD out - and got slapped with anticompetitive lawsuits. AMD has proven they've got solid designs with the PS4 and XB1. Only time will tell if Apple cares about name and reputation (such as it is) versus solid technology at a reasonable price. Any new shiny computer from Apple will be priced the same high price, regardless if it's Intel or AMD inside. Besides, Apple has been making noise about developing their own ARM chips for laptop computing. So this may all be a moot point.
The era of specialized hardware will return. We'll have sound cards, video cards, AI accelerators, etc.
I don't think so. Audio, video, and NN-AI are all doing the same thing: lots of low precision fp-ops in parallel. There is no reason to have separate hardware for each.
The way I understood it, Intel developed a ton of brand new techniques for getting 10nm and beyond going and it turns out attempting to use all of these new techniques at once on a new node resulted in it taking way, way longer to do and even when they got it sorta working the yields were awful.
It's not that difficult for Intel to prepare. They can buy chips from AMD. They can license ARM.
I'll see your senator, and I'll raise you two judges.
Seems he is misinformed or SemiAccurate is off base
https://www.cnbc.com/2018/10/2...
Just start developing some larger real-C++ codebase.
a majority of games are now written for tablets/phones that have little more processing power than a 486
*cough* Bullshit *cough*
All ARMv8 cores are between high 4's and mid 6's in DMIP/MHz and run at speeds from 1.5 to 2.5GHz, that puts them at 6750 to 16,000 DMIPs. The 486 DX4 100 was 70 MIPS and 0.7 MIPS/MHz. The modern ARM chips performance puts them in the range of an Athlon64 to an Athlon64 x2 chip.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
https://www.macrumors.com/2018...
AMD will likely have 7nm chips out sometime in late 2019.
It feels like it pretty much ended about 2010
I don't think it ended then as much as we have hit the point of diminishing returns.
We'll do a classic car analogy - lets assume that this is a compute controlled car that can be safely piloted at any speed stated.
Think about it this way: If you're going 5 Mph and you'll reach your destination in 96 hours, then doubling your speed is huge - at 10Mph you've saved 48 hours!
Then at 20 Mph you've saved 24 hours! And at 40 Mph you've saved 12 more hours! All that is great. Who wants to sit in a car for any longer then necessary?
The thing is, by the time you get to 81,920 Mph, if you double your speed again you're only cutting about 11 seconds off of your travel time. And it's only a few more cycles of doubling before we're getting into fractions of a second worth of difference. The increases in speed are no longer really relevant.
A while back - I'd say around 2010 or so - the home computer reached that point of just being "fast enough". People with specialized uses still want more speed, but its harder for specialized uses to drive market demand when general users aren't subsidizing the cost via volume anymore.
"People who think they know everything are very annoying to those of us who do."-Mark Twain
I'm not a semiconductor fab expert either, but I do know that some times big jumps like that are counter-intuitively easier.
For example, 10gbps network connections were starting to push the physical limits of a single wavelength on fiber and of the electronics behind the laser and photodiode, 25gbps is about as good as it gets with single wavelength direct detection on-off keying photonics. However, with 100gbps optics (real ones, not LR4 or SR10) it forced the need to use DSPs to generate the transmit and process the receive. The side effect is that suddenly the 100gbps optics have much higher noise and dispersion tolerance than the previous generation of 10gbps stuff. Now you can go out and buy a single wavelength 400gbps optic if you have the need and the budget.
Maybe the move to 7nm fabs will also involve a fundamental change in some other aspect of lithography that makes it much easier than 10nm was.
09 F9 11 02 9D 74 E3 5B D8 41 56 C5 63 56 88 C0
As far as I'm informed it's not AMD's 7nm process to begin with. They haven't had their own fabrication since 2008 when they chose GlobalFoundries. Apparently they took this step because of the costs involved in every subsequent node shrink. What they do now is to develop the CPU design and TSMC will be using their 7nm process to implement said design.
Theoretically Intel could do the same if they wanted to. Outsource at least some of their production until their own processes have become good enough.
The big question is what it would take for them to do so. Personally I think we won't be seeing this happen as long as enough people are willing to pay for their overpriced CPUs, just to get the best possible performance in video games.
Pentium III with RDRAM
Wasn't that supposed to be Pentium 4, or did I miss something about P III's back then?
Ezekiel 23:20
There was nothing specifically about who will do it, though.
Ezekiel 23:20
If you're running a VM for all current versions of Windows, at the same time, along with 10 Linux VM's just because, you might be a power user. Or maybe nuts.
If you're not trying to run all those all at once, you could easily get away with a pretty normal PC. Maybe the only slightly pricy bit could be a large enough SSD to hold all those VM's. Or just buy a hard drive for $75.
I don't think it's physics as much as financial. The demand is currently not there. A 5 year old laptop was faster than most current laptops/desktops/tablets/phones. Most people have stopped demanding high performance CPUs. Even gamers have shifted their focus to the video cards not the CPUs and a majority of games are now written for tablets/phones that have little more processing power than a 486. Because of the low processing power and low energy requirements of phones/tablets, the stuff that might have created demand like voice recognition has shifted to the cloud instead.
Moore's law was and has always been about economics. It is about the cost per transistor even at the expense of performance. Integration increases with greater density, larger area, and improved packaging which all contribute to lower cost per transistor.
https://www.youtube.com/watch?...