Slashdot Mirror
The End Of The Innovation Road for CMOS
Elledan writes "According to this EE Times article, CMOS technology (also used to create CPUs with) is getting near the moment when we will no longer be able to create smaller structures with it. With the date for this moment set around 2012 and with no replacement technology in sight, this issue might become a real problem in the near future, as the article explains."
At what point does the performance of computers become "adequate"? Once a technology becomes mature, a slow rate of improvement becomes acceptable. Reliability gets fixed, design improves, niche markets get filled. Internal combustion engines, houses, aircraft, ships, bridges, for all of these the lack of a Moores Law isn't a "problem". Perhaps if Moore's Law finally packs in for computers, we can all stop chasing progress and concentrate on things like social implications, human factors, and software that does something useful.
Panurge has posted for the last time. Thanks for the positive moderations.
How many times have we heard this prediction?
I remember when 200mhz was the end of the road. 'They' always manage
to give us another 10-15 years. It's like drilling for oil.
Besides, while Mhz makes a big difference to speed, design is more important.
Even if we hit this wall, we'd just continue to improve in other areas.
This is a different kind of FUD, but FUD it is.
rhadc
I say this is a good thing. Let the end of CMOS come. It's time for us to move forward. I think this is just the kick in the ass we need to really start focusing on quantum computing. IBM and Fujitsu both have quantum computing research divisions, and I wouldn't be surprised if there aren't quite a few companies out there very quietly working on it. The pressure for faster and better computing will drive us forward. And when the first 64-qubit computer comes rolling down the line, I'm certain Tom's Hardware will be there to tell us how many FPS's we'll be getting in Quake8 with it:
Tom's Hardware: I can definitely say that this thing smokes. Unfortunately, due to quantum uncertainty we weren't able to give you an exact measurement of FPS's. but we can say with some confidence that it's between 189 and Infinity + 2. However, with quad-sampling anti-aliasing on, don't be surprised to see that number drop to Infinity + 1.
Damn, I need to get some sleep.
---- El diablo esta en mis pantalones! Mire, mire!
MEMS isn't an electronic system like MOSFET or CMOS, it's a method for making mechanical systems out of silicon. Oops.
Would it really be so bad if manufacturing advancement in the hardware sector slowed or stopped? Companies would be forced to develop new features (like MMX or AltiVec) to differentiate their chips. Work would shift to bringing the rest of the computer up to the top speed of the processors, which it has lagged behind by orders of magnitude for years. The oft-hated hardware upgrade cycle would slow down greatly. Machines would be useful for much longer, and depreciate less. Software developers could no longer rely on increased performance, and would be forced to do real optimization.
I don't think anyone is suggesting that this is going to be the end of increased CPU speed, just the end of the usefulness of a certain technology.
I think perhaps the best thing that could happen would be about a five year freeze on increasing CPU power, so that the burden would again fall on the programmers to write good fast code.
In the past five years, CPUs have increased in speed tenfold, but computers have gained little apparent speed (applications don't load any quicker, OSes don't boot any faster) and certainly haven't gotten *ten times* more useful.
We have all these extra cycles, and all we can think to do with them is write slow, clunky but pretty window managers. (A criticism I lay against, MS, Apple, and OS) A pause in the mad rush for speed might give some time to think of what to *do* with all that power. DivX is a pretty specific use for so much general purpose hardware.
In Capitalist America, bank robs you!
A lot of people are saying that stopping/slowing hardware advances would improve software - it won't. The proof is in the gaming area - they optimise it as far as possible while still making the game profitable and they need the latest hardware all the time.
The reality is that software has aquired a whole heap of extra features that we take for granted and they take up space. There is no reason to want highly optimised code because it limits what is financially viable to develop. Optimisation is hard, takes time and costs a heck of a lot of money, there would be a lot less software out there if it had to be fully optimised to be usable.
In short, just because you think you're "l33t" by optimising your code so it runs on a 286, you're just wasting your time because computers are fast enough to not need that. Look at the amount of processor cycles donated to projects like Seti@Home - there is no need to optimise code, so stop whinging.
Slowing the hardware improvement cycle will just slow/stop the innovations in software. The first place it will hit is scientific areas, then the gaming arena and it will hit the average user because the cool high-end stuff just won't trickle down to them (like video editing has recently done).
10. To decrypt those files Mulder stole from the Pentagon.
9. John Connor has smashed your defense grid, and you need an edge, pronto.
8. Nothing can cheat like a quantum aimbot in Quake 4...
7. Negative ping times.
6. The shifty eyed salesmen at CompUSA talked you into it.
5. Opens up the exciting new possibility of quantum porn.
4. Windows.NET 2010 runs like a dog on your 2048-cpu, 900 Teraflops cluster with 8 petabits of ram.
3. The ability to render away the clothes, in real time, of your favorite TV show.
2. Your scheme to perform nuclear yield simulations with imported Playstation 2's ended in a trade embargo.
And the #1 reason to like quantum computing is...
*drum roll*
According to this paper (pdf) entitled "Scaling of Electronics" from 2001, the following conclusions are drawn:
* Moore's law will hold for 20 more years.
* There is a potential performance increase of 10000x with current CMOS-technology
* The minimum gate: needs 12(!) electrons to switch.
We'll see. I wouldn't hold my breath waiting for CMOS to hit the roof though.
Chip makers complain because the "CAD Community" isn't coming up with solutions to some of their problems, but University R&D programs are unable to keep up with fabrication standards as the equipment gets more expensive.
Isn't this a problem waiting for a few self-interested chip-makers to whip their wallets in the direction of a few universities?
But even the article repeatedly says that the mask cost issue is a problem for the little guys, not the large ones like Intel. They can and will cheerfully swallow $600k respin costs, and more, to tapeout a successful new processor. So this aspect won't hurt processor development at all.
As for the CMOS scaling issue, the processor companies- i.e. Intel and AMD, have the pockets AND the incentive to find work-arounds. I promise you all that processors will continue to advance well beyond the year 2012. It may not be CMOS, and it may not be pretty :-), but it will work.
So for all of you who posted asking "what will we do when processors no longer advance", let me set your mind at ease- THAT won't happen for a long while yet.
Finally, while the subject of my post is "the end of processor advancement", I'll say a few words regarding other types of chips. I work as a hardware engineer for an ASIC house, and we produce at TSMC using the 0.18u process. The point is, that for our chips there is NO incentive to go to 0.13u or below. Nor will there be a reason for quite a while. The same is more or less true for MANY MANY other ASIC companies. So while the cutting edge- processors, Flash and graphic-chips companies will probably need to switch from CMOS to some other technology around 2012, that will in no way spell the end of CMOS, not for a VERY large segment of the ASICs market, and not for a VERY long time.
I work in research at a university, and my PhD project is going to help solve this problem (and others) long before 2012. I can't get into specifics because of disclosure issues. But, understand that already a HUGE amount of work has been done behind the scenes and most other researchers don't yet know of what's to come.
.18um fabs could easily be refitted with strained Si material and compete with the .13um fabs. Actually, it's even better than that--the increases in mobility have been up to 8 times over that of Si.
CMOS isn't going to die. Turns out that we're not limited in the horizontal direction like everyone predicted years ago (remember how lithography was always the big problem?). Instead, it's the vertical direction. Our gates are having to get too thin. SiO2 just doesn't work well with 10A thick layers because of trapped charge and whatnot. Also we can't properly control doping at very shallow levels.
But all that doesn't matter. Strained-Si technology is where it's going. If you're interested, check out AmberWave. It turns out that we can increase the mobility of holes and electrons--so even older
No, CMOS isn't going to die. It's going to change and morph. Just like it has in the past. We don't need a revolution like many engineers are claiming--we simply need evolution. Strained Si is an evolution that will make for revolutions later. Current fabs can just swap out their current Si wafers and get strained Si ones--most everything else in the fab stays the same. Talk about a huge cost savings to boot (no need to rebuild a new fab for billions).
Long, cute, or funny Sigs are just another form of over compensation, used by geeks, nerdz, etc.
Here's a few quick explanations of some of the key points mentioned in the article.
The leakage problem: This is a really difficult and nasty problem. It arises from the fact that designing a chip involves trading off a number of things, among which are clock frequency, operating voltage and power dissipation. It turns out that as you increase voltage, it speeds things up but it also causes power dissipation to rise as well. Ask any overclocker. However, the speedup is roughly proportional to voltage, while the power dissipation goes as the square of voltage. Hence the operating voltage of chips has steadily been decreasing. The bleeding-edge research type chips are down somewhere in the 1V - 2V range. The problem here is that there is a fundamental property of the FET called the threshold voltage, the voltage at which (more or less) the transistor switches from being ON to OFF or vice versa. Of course it's not a sudden transition, so its desirable to have the system voltage higher (say by 2X to 5X) the threshold voltage, so that the transistors are turned ON and OFF fully. Otherwise, leakage occurs, and can become a very significant power drain if not kept in check. The problem is that due to physics and some other factors, the threshold voltage cannot be reduced easily past a certain point. There are tricks that the designer can use to attack this, but it's still a very fundamental issue. So what the circuit designers end up doing to meet the design criteria is play a large game of cost-benefit analysis with regards to power, frequency, system voltage, threshold voltage, area (die size), etc.
Masks: Integrated circuits are build up in layers. An extremely simple design might have 6 layers, modern CPUs might have 20 or more layers. Each layer is created with a mask that defines the features of the layer. While enlargement/reduction is used (meaning the mask features are larger than the features on the wafer), mask creation is still very difficult. It's like making a stencil with millions of tiny features. The photolithography involves very expensive machines with extremely precise optics. Indeed you might have heard of the push to "extreme ultraviolet" - this refers to the light source which shines through the mask and exposes features on the silicon wafer. The trend is to use smaller and smaller wavelengths, because the feature size keeps shrinking. The wavelength of light that is used must be significantly smaller than the smallest feature, otherwise you get interference/fringing/etc. Anyway, these masks are very expensive to produce, leading to very little room for error. You want to be sure that those masks are at least functional, and hopefully as bugfree as possible. To a certain extent you can work around some hardware bugs, but it's very stressful because of the huge cost and time delay (many months) of getting a design fabricated. Imagine what development would be like if compiling your source code one time cost you a million dollars and took 6 months. Now try to stay competitive in a market where everybody is screaming at you to get a product to market as quickly as is humanly possible. Simulation is the name of the game here.
Interconnects: This refers to connecting together the individual transistors to form blocks, connecting the blocks to form modules, etc, up higher and higher levels. Interconnects do not scale well, it's just one of those complexity things. The number of interconnects goes something like N^2 (where N is the number of transistors), and this can quickly get out of hand. The problem is you can't just make the wires longer (by wires I mean the etched paths inside the chip, not the external things) because this increases their resistance and capacitance, which means that they must be driven "harder" to achieve a given performance. To drive them harder you must spend extra area on larger transistors (which just complicates things -- now the chip is even more spread out) or spend more power, which is usually not feasible. A stopgap measure is to use copper instead of the traditional aluminum for the interconnects, but this is only really a one-shot thing, it only buys you so much. Another way is to use more interconnect layers (expand in the "z" direction) but this has its problems as well. The most promising solution to the interconnect issue is with advanced CAD algorithms and plain old good design. Keep related modules close to each other, and design busses to shuttle things around longer distances.
Capacitance: Capacitance is one of the worst enemies of the circuit designer. It means that on every transition of state, energy must be spent charging (or discharging) a dielectric. This is one of the main reasons for reducing smaller feature size -- smaller things have less capacitance. The article mentions fully depleted SOI, which is basically a very extreme way of trying to reduce capacitance. The bulk substrate is silicon dioxide, an insulator, instead of pure crystalline silicone (a conductor.) The effect is to decouple the individual transistors from the bulk substrate of the wafer. The result is much less stray capacitance, but the cost is that your transistors no longer work quite right so it makes circuit design that much more complicated. The article also mentions high-k dielectrics, which basically is a way of increasing the "gain" or drive strength of a transistor without increasing its size, which is the normal way of doing things. It can be really quite frustrating: if a path in your circuit is too slow, you have to increase its drive strength. But this also increases the capacitance (which leads to more power dissipation) and now the thing that drives that circuit also has to be bigger (to compensate for the increased gate area), etc, etc. Any means of increasing the drive strength without increasing area is quite beneficial.
I hope that was of some use to at least someone.
Historically, increased CMOS speeds have come from one thing: shrink the features. Atoms being small, this works for quite some number of doublings. Techniques such as strained Si, alternative gate dielectrics, etc. are a qualitative change in strategy. They have potential to help, but they don't have the long-term extendability that we've seen from shrinkage. Let's say strained Si gives a factor of 8 in mobility. That's great, but in 3-4 years it's done and we need some other idea orthogonal to the previous one. Having to come up with a qualitatively new enhancement every 3 years is very different from the make-it-smaller world to date.
Curtains for windows?
Remember the good old days when a good engineer could race a computer to a solution with a circular slide rule? I do. Then there were complete IC based computers and we couldn't do that anymore. Then around 1987 we all said 25 nano lithography was the theoretical limit of the physics. Which of course was wrong because it was based on materials science that was already old.
At any rate - I don't feel comfortable making prognostications about technology 10 years in the future. Any every time I think about I also think about Turing's paraodx. That says, that if you need 10 years to solve a problem today but in 3 years you will probably have the technology to solve it in only 5 years then you should wait 3 years to start and you will be 2 years ahead of the games already.