Slashdot Mirror
Intel's 14nm Broadwell Delayed Because of Low Yield
judgecorp writes "Intel has put back the delivery of its 14nm Broadwell desktop chip by a quarter because of a manufacturing issue that leaves it with too high a density of defects. The problem has been fixed, says CEO Brian Krzanich, who says, 'This happens sometimes in development phases.'"
The good news is that it is just a defect density issue. A first round of tweaks failed to increase yield, but Intel seems to think a few more improvements to the 14nm process will result in acceptable yield.
14 nanometers should be enough for anyone.
Upto 50nm, it was fine, but is now in the region of diminishing returns. The cost savings that were always synonymous w/ shrinks are no longer there, since the process costs easily outweigh the cost savings per die, even assuming a 100% yield.
Because a wafer of pure silicon has a high, fixed cost. The smaller you can make a chip made from that wafer, the more chips you can make and the lower the cost for each chip. Intel's recent 22nm chips have recently been around 200mm^2. If they were made on 640nm, they would be about 200*640/22=5800mm^2, or a square of about 7.6cm (3 inches) on a side.
Also, smaller wires use less electricity and generate less heat. You wouldn't want 640nm Haswell in a laptop.
(T>t && O(n)--) == sqrt(666)
Potentially it can keep going until the size of a transistor is just a few electrons across but as we get closer to that point quantum teleportation becomes more of an issue. This is cool video that explain some basic stuff about transistors and the end of Moore's Law.
https://www.youtube.com/watch?v=rtI5wRyHpTg
For others like me who had to look this up, link.
You had to look this up? Now I feel old. Also, get off my lawn!
Maybe we could build computers out of Planck planks. They're really small.
I should use this sig to advertise my book ISBN-13 : 978-1501515132.
Intel has produced two new generations of processor that were WORSE than Sandybridge. Higher power use (under load) and far less over-clockability. The newer part were ONLY better (in desktop systems) if you intended to use their new instructions (vanishingly unlikely) or the integrated graphics (which would be pointless- people buy expensive Intel CPUs to partner them with expensive GPUs from AMD or Nvidia).
Intel, of course, were in the same position with the waves of Core2 parts, each of which essentially overlapped each other in performance generation on generation (although power consumption was much improved over the first generation of Core2 i7 parts).
Intel currently doesn't know exactly where to go in the near future, and is attempting to hedge its bets by trying various things. It is currently undercutting its own HYPER-expensive ULV mobile high-end parts with the new 4-core 'atom' Bay-trail chips that seek to go head-to-head with ARM. Because current high-end ARM is so good, Intel is forced to sell a very dangerously good chip (dangerous to Intel's profits, that is) into low and mid-end tablets, running Android or Windows8.1
However, even Intel's first decent 'Atom' part ever (after 5+ attempts) is beaten by Nvidia's somewhat lame Tegra 4, and Qualcomm's Snapdragon 800. It is exterminated by Apple's new ARM chip, soon to be seen in Apple's new iPad refresh.
Intel's 22nm process, and use of FinFETs, has been a total disaster so far. A process advantage, and custom designed chips, doesn't allow Intel to beat ARM parts coming from commodity foundries at TSMC and Samsung in 28nm. Sure, Intel can make its own chips smaller than those on the previous process, and theoretically get more parts per wafer, but the per wafer costs rocket, the yields drop (initially), and insanely expensive new plants have to be built to service the new process.
What does Intel get from spending all this new money on R+D? At this moment in computer history, almost nothing. The x86 is dying, and everyone BUT Intel builds ARM solutions. Every major player has a GPU (graphics) solution as good as Intel, and Intel isn't within a million miles of matching the AAA-gaming GPU designs from AMD and Nvidia (despite the fact that Intel has spent more money than every graphics company combined, across their combined periods of existence, to create its own GPU solutions).
Intel simply has no current use for its expensive 14nm process. It has built the factories, so it is engaged in a waiting game- waiting for mobile parts to roll off the 14nm production lines that have clear market advantages over its current mobile chips. It just isn't worth Intel's time launching another round of non-improved parts. The market has changed forever, and on-one wants to buy "this season's Intel" for the brand loyalty reasons previously apparent.
Intel fanboys want 6-core and 8-core parts, but Intel is extremely loathe to risk introducing better value into the desktop market. If Intel properly sold 6-core solutions, they would have to sell 6-core i5 parts, and these would beat-up their EXTREMELY profitable 4-core i7 parts. Intel is too in love with the status quo.
If Intel's bay-trail 4-core parts prove good enough for tablets and non-gaming laptops, and they will do having greater performance than the more than adequate mobile 2-core core2 parts used in the first decent cheap laptops years ago, where does most of Intel's mobile biz go from here? Bay-trail parts (unlike those years old mobile 2-core core1/core2 laptop chips) also do all the video decoding in hardware, allowing flawless playback of all current video content (and bay-trail is strong enough to do CPU enhanced decode of 4K video recorded in h264).
Bay-trail is the part Intel moved Heaven and Earth NOT to produce. Bay-trail is the final step on the race-to-the-bottom for x86 based computers that most non-AAA gamers will need. If the only real money Intel makes ends up from chips lie Bay-trail, Intel is done.
Think about this. In a few weeks, you
The main reason die shrinks happen is usually at the behest of the manufacturer, and rarely at the behest of the customer (unless we are talking die sales, which we usually don't for Intel, and which is a whole different ball game). There are always downwards pricing pressure on the manufacturers (a tad less on Intel, I'd think, given their elite position amongst fabs) and in order to preserve their margins, they work in these shrink transitions w/ their customers. But make no mistake - for customers, those shrinks imply requalificaiton and a whole new product development cycle before they can go to market w/ those. They'd rather get their price cuts on the same die, except that the manufacturers typically won't give them that.
In the past, the reason to go for shrinks was improved clock speed, and more recently, it's power consumption. But power consumption alone doesn't drive such market trends, particularly given the expenses incurred - what really drives it is cost. But again, as I said, we're really past the point where shrinks would result in any significant cost savings.
So make a "triple core" edition called an i4 where really 1 core just didn't pass quality control so they turn it off. AMD did it and it sold so well they had to purposely cripple working quad cores to meet the demand for triple core chips.
I have a theory that those T-edition chips Intel made that are just underclocked, hyper-efficient, ultra-low wattage editions of their recent chips are actually just ones that wouldn't run properly at the normal stock clock. I never heard a solid claim that they actually had different voltage regulation circuits or something like that. They just underclocked them and made them have a higher tendency to not click to a full multiplier level as often or for as long.
Might explain why I went with an AMD processor, and might get a video card from them soon.
'I like my chips like my women: hot and slow.'
TSMC's metal 1 pitch is 64nm in 20nm, and Intel's 22nm is 90nm.
14/16 is indeed expected to have ~64nm pitch, so it's not better than TSMC's 20, but it's a great leap for Intel.
"disabled": Not quite. When I'm running 491MHz internal, you can't just disable arbitrary logic on me. The slower parts may get away with disabling columns, but you can't change my timing without breaking my design.
Also, all the hard IP is not redundant, and there's more and more of it.