'We will not rest until the periodic table is exhausted' says Intel CEO on quest to keep Moore's Law alive Intel won't give up on Moore's Law, even though just about everyone has declared it dead or on its way out. "Moore's law is alive and well," said CEO Pat Gelsinger during a keynote at the Intel Innovation event, which was webcast on Wednesday. He showed a chart tracking the semiconductor giant progressing along a trend line …
I foresee a couple of problems.. First, as the transistor count escalates and feature geometries get smaller, the chance of device failure inevitably rises, as it can take only one transistor to burn out to degrade or disable the device. Second, as features get smaller, more numerous and more tightly packed, the chance of functional errors gets higher, not only due to externally sourced interference but as a result of internal crosstalk (particularly as clocking speeds also rise). We may eventually (soon?) reach a tipping point beyond which it's impractical to go if we want reliable devices.
The reason clock speed hasn't increased in the last decade is because, while Moore's Law continues to hold, Dennard Scaling does not. Dennard Scaling is the notion that energy use is roughly proportional to chip area, so as transistors shrink, you can fit more in the same area without using more power -- or you can reduce the chip area to a quarter and double the clock rate with the same power usage (power usage is roughly proportional to the square of the clock rate). But Dennard scaling stopped around 2006, so to get higher clock rate, you needed to increase power use (and heat dissipation).
As a consequence, the trend shifted from increasing clock rate (with quadratic power usage) to having more cores per chip (with linear power usage) and shutting down cores (and functional units) not currently in use. This is why your laptop fan goes crazy under heavy load -- all cores are in use.
So the problem is not really shrinking -- that gives relatively little. The problem is power use, and that has somewhat been addressed with FinFETs and specialised transistors, but that will only take you so far. Other materials may help, as will superconducting transistors (but these are fiendishly difficult to combine into a complex circuit, as they interfere heavily with each other). Other potential solutions are asynchronous processors (so you don't have a global clock) and reversible gates (which have a lower theoretical power usage than traditional irreversible gates such as NAND and NOR). But these have yet to be realised on large scale.
"what you said" plus lowering the voltage way below the 5V these systems used back in the 90's, so that you can use more current for switching, switch less voltage (making it a bit faster), and NOT burn out everything from the high wattage.
And, using different gate, channel, and substrate materials might help run on even LOWER voltages.
schotkey diodes have a forward bias voltage similar to germanium (0.2 to 0.3v) while silicon needs 0.6 to 0.7 . different dopings and materials might improve this enough to run at under 1V, let's say, allowing currents of ~50% more than what we're using now.
i think the power supply on my newer Ryzen boxes is in the neighborhood of 1.5V or something like that. Not sure how long they've been like this though.
I agree with your general point (Dennard has been more important than Moore). But. I think it’s deceptive that the “top of tree” problem for CPUs is power-usage, rather than gate-delay (clock-speed).
Imagine you could wave a magic wand tomorrow, and reduce core logic power consumption by 10x. Could you just up the clock by 10x? Not really. The core would just starve due to the memory wall. Nor does hyper-threading get you out of that - you still run out of on-chip cache. The economically allowable gate area of the *cache* sits at most a factor two underneath power-usage in determining clock-frequency. A magic wand to scale the cache gate-area by 10x, actually could probably scale performance by maybe 3x without breaking power-usage constraint. So Moore is important.
They are related, though. There are 3 factors to CPU power consumption, dynamic consumption, short-circuit consumption, and loss due to transistor leakage current. Dynamic consumption is a function of switched load capacitance, voltage, and clock frequency, so is directly related to clock frequency. Short-circuit consumption is also dependent on clock frequency. Leakage current consumption is dependent on supply voltage.
Yes. The PowerVia tech is a power plane below the transistors that allows both direct power to the transistor (through a via below the transistor) and essentially eliminates a voltage drop. It also frees up space in the interconnect stack above the transistors to allow for relatively larger interconnects to reduce resistance and capacitance between lines. The result is the ability to switch faster, so we may once again see a more significant bump in clock speed with 20A.
Transistor failure is an issue now, the solution implement by everyone is to build redundant circuits that automatically take over when the primary circuit fails.
A much bigger problem that they are still trying to work out is how to compensate for, or even exploit, quantum effects that start to occur when you try to build transistors that small.
Cpu failure is not what makes a device unusable at the moment and I don't forsee that in the near future.
Density per _device_ is easy to double by chucking in another whole chip, or a task specific chip. You could double the volume of a modern phone or laptop easily by making it fatter. It's not really needed and expensive but it seems to me its far from infeasible.
I had a really good chat with one of Intel's designers a few years back, and I raised a similar point.
He said that there's a degree of redundancy built into the architecture so a single transistor blowing won't affect things too much, and also that they've been correcting for effects like quantum interference for a while.
Again, Arizona. The driest state of the Union is where everyone is building in an industry that needs millions of gallons of water every day.
Is there some special virus in Arizona that systematically infects only semiconductor CEOs and bends their mind to choose Arizona, instead of, gosh, just about anywhere else that actually has water ?
Arizona has a relatively low cost of living (not run by a certain political party like Cali-Fornicate-You).
It also has a LOT of open land. They also have less expensive electrical power, apparently not being mandated to use "alternative" energy at much higher cost.
In short, LOWER OPERATING COST. And being on the border of Mexico, next to California and Nevada, and one state away from Texas, it's actually well located. And for various reasons the airport in Phoenix is a MAJOR connection hub.
There are other advantages to Arizona as well, but last time i was there southern Arizona was ok (close to Mexico, along highway 8) but in the north my allergies went berzerk.
whether there are tax breaks or not I do not know. But the dry climate might actually do better for making chips, too. Just a thought.
shipping things around by rail from Arizona would be pretty economical as well. It's way south of the Rocky Mountains and Sierra Nevada mountains, two major hurdles for railroads.
The counterargument is that the water can be recycled - up to a point - but some water is still required.
The problem will be if Arizona's water supply decreases, and temperatures rise, even further. There is not even an ocean nearby to run desalination plants.
For the gushing review of their polished mirrors and atmospheric smoke. RibbonFET might be a game changer, but was not demonstrated, and might not work.
The reality is that Intel need to "promise" jam tomorrow because it has a weak position, and about to take a beating. AMD make faster AMD64 chips (x86-64, x64,.. are aliases), nVidia makes more scalable chips, and Apple is humiliating them with M1 max chip.
One of the main reasons Apple turned to its own Apple Silicon SoC's is that Intel was not delivering what it said it would. Apple designed most of its laptops based on the specs Intel gave them and then Apple found they ran hot and got a reputation for throttling. I also suspect that Apple had Apple Silicon in mind for computers from the get-go after putting it in iPads and iPhones in 2012.
Intel has obviously taken its eye off the ball and its going to take a lot of work for it to catch up.
“Apple is humiliating them.”…..
No….Apple has *shown* them.
To a man with a hammer everything looks like a nail. To a CPU manufacturer with a generational history of making chips to a totally standardised memory architecture, there is only one way to do things.
Apple, coming from a smartphone background, realised that SoC architecture plus soldered in memory SIP was just going to destroy the conventional architecture. It really is largely the memory architecture that makes the M1 fast. Having a power-efficient *implementation* of ARM, also omitting the legacy cruft of forty years of backwards-compatible CPU development, is second.Arguing about the ISA (x86 vs RISC) is last century’s war. Makes really only a few percent difference.
Within a single CPU generation, Intel could take one of their cores and re-cast it into a SoC architecture with HBM. And then they will be competitive again. The only thing stopping them is stubbornness. And unless they are idiots, they are doing just that, under wraps.
Good luck to them when it comes to using the heavier actinides and above. The longest lasting isotope of Lawrencium has a half life of 10 hours, for Copernicium it's down to 28 seconds. Not something to keep on a warehouse shelf.
"Transuranic heavy elements may not be used where there is life. Medium atomic weights are available: Gold, Lead, Copper, Jet, Diamond, Radium, Sapphire, Silver and Steel. Sapphire and Steel have been assigned."
WTF (ium)?
Jet?
Diamond?
Steel?
Sapphire?
Somebody is demonstrating that their degree is in marketing...
I just have to wonder if, when they get to using francium (22 minute half-life), the hardware replacement cycle can really compete with the software-forced replacement schedule.
Would it even be safe to be in the same county as a 30cm wafer of Francium?
(not skilled or curious enough to actually pursue an answer)
that what Apple has created, will no doubt have other companies (Samsung, Huawei spring to mind and maybe that company that makes the world's modems whose name I cannot remember)) thinking if Apple can do this, so can we.
Perhaps the curtain is starting to close on the whole x86 thing?
Chipzilla will do themselves a favor if they just fire the moron. Moore's Law is an observation that with time key computer hardware is/was getting faster as we got better at making denser integrated circuits. Since it was an observation, which has been revised a couple times, it will eventually fail as it was not based on any fundamental physics that will allow unending speed increases. It was observation about how we are getting better at designing, engineering, and manufacturing faster chips. How much longer will 'Moore's Law' still work, not sure, but we will eventually run into a brick wall. It does not matter if you find another semiconductor instead of silicon (germanium has used historically and I think some dabbling with gallium arsenide has been done) or not there is a brick wall according to our understanding of physics.
It was quite a few years back that dynamic memory got down to the round dozen of electrons to store a bit state. It got difficult to go below this because you need a crowd to form a consensus -- you go too low and quantum effects start introducing significant numbers of errors. Its the same with processor geometry; I'm not sure just how low people can go but you hit diminishing returns, incremental improvements costing increasing amounts of effort.
The solution to not having enough speed has been to go wide. This is why our machine words have been increasing -- its not that we have to use 64 bit or even 128 bit integers for our everyday tasks but its a simple way to double or quadruple memory speed. What we've been less good at is parallelism -- even though our processors now boast multiple cores they're really just multiple single cores; we really haven't figured out how to multiprocess unless the tasks lend themselves to or can be adjusted to fit that architecture. (The nearest I've see to multiprocessor operation in a normal program is register renaming in a RISC where the processor effectively runs two execution threads based on the future outcome of a jump decision, the unwanted one being discarded once the jump is known. Ungodly complicated, ingenious but really only scratching the surface.)
Anyway, if he thinks Intel's got some secret sauce that will magically make things work faster, good luck to him. But he probably needs to learn a bit about trade offs.... (universal law of the Universe -- "There Is No Such Thing As a Free Lunch").
Intel isn't even using one EUV (or equivalent) machine yet. (Although they have one next gen EUV pre-ordered from ASML for ~2024. )
How did they miss that bandwagon? Insisting on in-house vertical integration?
I see no explicit recognition of that strategic miss. Therefore the risk continues.
Moore's law isn't the problem - the problem is not keeping up with competitors.
I wish Mr Gelsinger Moore luck at catching up with the competition.
At least he is focusing on fabrication and prepared to invest in it - an essential prerequisite but not a guarantee of success.
His focus is on distracting attention from the facts. There isn't anything else he can do, other than nothing at all. 7nm still doesn't work. Everyone else is moving on to 5nm. Intel's products on the market still aren't competitive. Fixing these problems takes years; it's not something he could do overnight even if he knows literally everything about the fabrication process. They're hoping they can fix all those problems before AMD gets production ramped up or ARM or RISC-V comes along and makes amd64 irrelevant.
In the meantime, Mr Gelsinger's job is to spew words and distract you. Is it working? Has this slideware and bosh about the periodic table and throwbacks to Intel's much more glorious past and speculative handwaving about what they might build next year or 10 years from now made you want to go out and overpay for a hot, slow processor? Have they excited you about upgrading next year to a new hot, slow processor that will also be late and even more expensive and still slower than what AMD shipped in 2019? Does calling 10nm "Intel 7" make you want it more? Does it seem less obsolete if they make up a number to write on yesterday's technology? That stuff's the here and now; would *you* want to talk about it, in his position? Of course not. You'd assemble a bunch of nonsense and hope some reporters will bite.
What's really sad is that El Reg continues to give them not merely air time but fawning, gushing, gullible as all hell air time. The emperor has been naked for a long time, folks, and the fact that their new wonder boy CEO will say practically anything, however idiotic and irrelevant, to avoid talking about Intel's actual products and capabilities says everything you need to know.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
