Slashdot Mirror
Intel's Tick-Tock Cycle Skips a Beat
New submitter Ramze writes: Several outlets are reporting on Intel's confirmation that it will make three generations of 14nm processors, delaying the switch to 10nm. The planned 14nm Kaby Lake processor marks the first time Intel has skipped the "tick" of a die shrink on its regular "tick/tock" cycle. Production of Cannonlake processors on 10nm has been pushed back to the second half of 2017 — likely due to manufacturing difficulties. Intel reported earlier this year that it may have to switch away from silicon to exotic materials such as indium gallium arsenide to make the next shrink to 7nm.
Hi there, young entrepreneur! Welcome to the Indium Gallium Arsenide Valley! :)
I know - the transistor count should have enabled us to build neural nets to filter out inane AC comments by now.
My God, it's Full of Source!
OUTSIDE_IP=$(dig +short my.ip @outsideip.net)
Look at Intel's recent earning and revenues. Business is so bad it doesn't justify investing money in a new engineering shrink.
love is just extroverted narcissism
This just in from Intel Markteting:
"Vee have vays of making you tock!"
Solving Unix problems since 1989...
Sodium and Chlorine being dumped on our roads when it snows? The horror.
Only the State obtains its revenue by coercion. - Murray Rothbard
Not to mention dihydrogen monoxide - which is still used in large quantities in many residential and commercial settings, despite its many dangers.
Why build neural networks when you can get the wild-type ones to do the same work in exchange for 'mod points'? Markedly more cost effective.
It certainly doesn't make the task compact or cheap(and appears to make it much harder to directly duplicate with either transistors or the inspired-by-but-not-models-of version of 'neurons' that 'neural network' usually implies in the context of computers); but the one nice thing about the very low 'clock rate' of the brain is that it suggests that however it does what it does; it can tolerate suffering fairly high latency per unit of distance between elements.
If you can tolerate latency, you at least have the option of buying more racks to compensate for what you can't achieve with available miniaturization and integration. If you can't, you run into relatively painful constraints on size. At the speed of light your hypothetical 3GHz processor is going to be waiting ~10clocks/meter for anything it needs from elsewhere in the system; and any practically available arrangement is going to be slower than that. If the brain were dependent on low latency, inability to replicate its density would(at best) mean being limited to replicating it slower than real time, possibly much slower. If it is relatively tolerant, replicating it at lower density might be horrifically expensive; but at least potentially doable with the ability to fabricate only rather small chunks of very high density.
Given that LEDs already lean pretty enthusiastically on those elements, without too much comment; and some of the stuff they bake into the silicon is pretty dreadful I imagine that they'll diplomatically ignore the issue and hope really hard that the Indium content is high enough to make the semiconductors that aren't super small and embedded in epoxy economically recyclable.
Getting changes made when there are at least partial alternatives and the material is distributed in substantial quantities throughout the entire product is much less of an uphill fight than getting changes made to the relatively small chunks of the product where much of the value and the profit are; and the producers say they've exhausted the options. They would probably be loathe to push it unless their position were markedly stronger than it appears.
Intel is stalling at 14nm. Everyone else stalled at 28nm. 28nm is still the cheapest node in per transistor terms. Since most chip makers are driven by cost rather than transistor performance, there have been few takers for 20nm and 14nm.
Broadwell was delayed by a quarter due to defects, not a year. 3 months is a mere glitch in the usual 12 to 18 months between Ticks and Tocks. This new delay pushes a regular part of the cycle out a year or more to when we'd expect another "Tick" -- effectively skipping that part of the cycle entirely.
If one were to say that they're "Adding a new tick, not skipping a tock", then we'd expect the usual shrink every 24 to 36 months with 2 architecture improvements instead of 1 in-between the shrinks.
As an analogy: If your employer gives you a $500 bonus on even months and a $20 Subway Gift Card on odd months, and then tells you that your usual expected $500 for August will instead be replaced with another $20 gift card, but that the $500 bonuses will resume in September and the alternating cycle will likely begin again, your company effectively skipped giving you a bonus. Your income for that year will be $500 less ($480 if you count the additional gift card as income).
If instead, your employer who regularly gives you $500 bonuses and $20 gift cards on alternate months decided to give you an additional gift card one month between regular $500 bonuses, THEN you could say that they ADDED a gift card. (or a tock in this analogy).
We're not talking semantics. The tick/tock is all about producing something on a regular schedule. You know, like clockwork. A die shrink is overdue and it's being replaced with an architecture change. The die shrink was planned at this time, and now is not. It has been skipped, so we go to the next part of the cycle which is an architecture change.
Skipping half a cycle and adding an additional part of the cycle only look the same if one takes out the time factor. It's the frequency that matters. You admit to this part of the cycle being "pushed out a year" (or more) when half a cycle length is roughly a year to year and a half. If you look at the frequency of releases as a wave with Tick as peak and Tock as trough, the wave function has skipped a peak and flat-lined at a trough to begin again. On a heart monitor, this would look exactly like skipping a beat.
Equivalent Xeon's get roughly a 2x multiplier for cost, as do the motherboards for them. My work machine is a 6 core Xeon (E5-1650 v2) that can be bought for about $650 compared to about $330 for an i7-4790k, which is also what is roughly expected for the i7-6700k when it arrives with its piddly little 4 cores sitting next to a vast wasteland of third rate GPU.
So either I would like a cheaper i7 without an on-die GPU, or more cores and cache in an i7 in place of the GPU.
With AMD continuing to gasp for life, I think it is fair to ask questions about what the hell the dominant monopoly in town is doing and why they seem to be stalled out. I am not in favor of breaking them up, but it is appropriate to scrutinize and control the pricing and behavior of companies that are in a monopoly position.
Performance per watt is wonderful and all, but we have had a lot of years where only the denominator has made significant improvements while clock speed and throughput per cycle are excruciatingly stagnant. I am complaining because I want a faster fricking machine, and it appears that Intel has either by willfulness or ineptitude has failed to deliver better speed in a any meaningful way in the last several years.
We are instead getting integrated crappy GPU's in flagship processors that will mostly never get utilized
Actually I think you'll find Intel is the number one graphics vendor in the world. The number of business machines with dedicated video card sits so close to zero that it disappears in a rounding error. They are also the largest market for computers in the world. That's not even taking into account low end laptops, sometimes high end laptops, tablets, and most low-power devices on the market.
Better on-board / integrated graphics processors have been sought after for a long time. You won't run Crysis on it, but then very few people have the desire to.