Slashdot Mirror
Intel: Metal in Future Chips = Less Leakage (updated)
securitas writes "Intel is contemplating using metal instead of silicon in future chips for the 'transistor gate, which controls whether a transistor is on or off' and the 'dielectric, an insulating layer below the gate,' which are respectively made of silicon atoms and silicon dioxide. 'Millions of minuscule switches that make up silicon chips leak electricity when they're supposed to be shut off. To compensate, engineers have increased the current, driving up power consumption, decreasing battery life for portable devices and generating more heat.' AMD has also experimented with metal instead of silicon. By moving to metal AMD and Intel expect to reduce electricity leakage. More from AP via SeattlePI and the Miami Herald." Update: 11/05 15:25 GMT by T : Read on below for some information from Intel on why this is a good thing.
gManZboy writes "Following up on the Intel news that about using metal in chips -- here's an explanation from Shekhar Borkar (Intel Research Fellow) about why heat, power, and sub-threshold leakage, not transistor size, are the real challenges to Moore's law. Apparently, in order to make chips much faster, we're going to have to pump more electricity in then anything else in our houses -- and they'll soon be as hot as a nuclear reactor -- no, really."
I dont see any mention of the type of metal that would be most suitable. I'm sure all metals are n't created equal.
The changes are largely necessary because of the unsavory consequences of Moore's Law, the famous dictum that states that the number of transistors on a chip doubles every two years. Yeah, it's all that pesky "Moore's Law" fault...
Man, and here I thought silicon felt weird.
I swear I remember IBM moving to copper for chips a while back (C.2-3 years ago). Was it for production chips or just R&D purposes?
Is this just a question of Intel playing catch-up?
Yeah, I hate it when my silicon breaks and creates leakage.
I have no trouble understanding a switch from poly to metal for gate connections... but a metal dielectric? That seems to run counter to common sense. The dielectric is, by definition, required to be an insulator, whereas metals, also by definition, conduct electricity rather well. What is this magic substance?
:-)
I love this site sometimes - where else can you post completely clueless questions and be virtually guaranteed to get an intelligent response from at least two people with PhDs in semiconductor physics?
These sigs are more interesting tha
to MetalValley!
:)
Now, instead of "experiment in silico", it would be "in metal" (??) or "in Fe|Au|Cu"
DNA in your Linux: DNALinux
...diamonds?
I thought that the manufacture of diamonds was set, and only needed to step up its production. Gemesis has been making, for less than $100, gems that would be worth hundreds of thousands if naturally mined.
The most promising thing about these diamonds is that, being cheap, they open the door for cpu cooling. Diamonds are tolerant of exponentially higher temperatures than silicon, so why aren't we hearing about intel, amd, motorola, ibm, TI, and sgi taking advantage of this new technology.
Metal? What about metal is unprecedented? What about it has kept us from using it before? Diamonds are the future, not metal.
Now how can you say that CPUs are based off of alien technology when Intel is making changes like this?
They just caught a new flying saucer.
45 5F E1 04 22 CA 29 C4 93 3F 95 05 2B 79 2A B2
The history of Moore's Law.
Or if you are interested in Moore's original paper, you can find it here.
I want to drag this out as long as possible. Bring me my protractor.
The chemistry of the non-silica gate dielectric requires that the gate itself be non-silicon, and metals are better conductors anyway. (For larger transistors, we're already running into trouble from the distributed resistance of the gates.)
Hope that helps.
Lacking <sarcasm> tags,
I thought it was already established that silicon implants were prone to leakage.
But switching to metal? Man, I'd hate to walk outside on a cold Montana morning in February with those.
What's that? Silicone? They're not the same? Never mind. Carry on. Sorry.
So you're telling me SOI is NOT a busty gal in an angora sweater?
Interesting how IBM has discovered that moving to metal for processors and away from metal for hard drives. (Newest Hitachi/IBM notebook drives use Pixie dust which is actually glass. The platters in these hard drives are also ferro impregnated glass platters)
Yell & scream & rant & rave... it's no use... you need a shaaaave ~ Bugs Bunny
Why is VLIW not more popular? Because compiler technology isn't yet good enough and current VLIW designs have restrictions that get in the way of the best performance.
Over the years, there have been many attempts to use techniques such as VLIW, which sound great on paper, but don't do well in practice. What have worked the best, at least through the 90s, are architectures that do a lot of simple things fast.
You can make VLIW fast, Intel has managed that, but at great cost in both silicon and software.
Be careful when making generalizations about a processor line such as the P4 - there have been quite a few P4 generations, each better than the last. Latencies have gone down.
I think that parallelism (eg. HyperThreading, multicore, etc.) is where the real-world performance gains will come from. Single-threaded benchmarks don't accurately reflect realistic workloads.
Here is an article explaining low-k dielectric. I believe this is a shipping product on the Power4/4+ based systems and it is in the EXA chipset on the x365/x440/x445/x450 Intel servers, and the Apple G3 and G5. The xSeries products even have little copper BB's in the grill of the system to symbolize that they use copper based technology.
As a rock-in-roll Physicist once said, No matter where you go, there you are.
My course in VLSI design was many, many years in the past, but what I do remember is that early integrated circuits used metal gates in the fabrication process. That process was later abandoned in favor of polysilicon because poly was much easier to work with at smaller feature sizes (I'm a bit foggy on this one). Gee, so now we're going back to metal gate processes, and we'll have real metal-oxide-semiconductor field effect transistors again?
If this is becoming easier to do at deep submicron level, I suppose processes for making deep submicron feature-sized Gallium-Arsenide MESFET's also got easier? Now wouldn't we just love to have such GaAs chips on our desktops... (I do know I'm forgetting another difficulty in working with GaAs, anyone care to remind me why GaAs is not as common as silicon today?)
Qu'on me donne six lignes écrites de la main du plus honnête homme, j'y trouverai de quoi le faire pendre.
No. The bonds between silicon atoms are covalent. A metal (e.g. copper) has a "cloud" of electrons free to move around in the lattice. Silicon is a semiconductor, with the charges bound to the atoms except when there's enough energy (typically thermal) to kick them loose.
Lacking <sarcasm> tags,
Would metal really be able to replace silicon? IANAEE, but...
Wait, that only works on the law forums. Darn.
One day, your computer may be the ONLY thing in your house connected to the outside mains supply!
TWW
"Encyclopedia" is to "Wikipedia" what "Library" is to "Some people at a bus stop"
As for the phonon question: in crystals the quantum of atomic motion is called a phonon. Electrons can scatter off of phonons, reducing the mobility and hence increasing the resistance. Two ways around this: lower the temperature, which suppresses the creation of phonons, or use a heavier material, which is harder to move and hence phonons take more energy to create. Using a metal gate dielectric (heavier material) traps any phonon which touches it, reducing the concentration of phonons near the surface, where the conduction electrons are.
Has anyone actually checked the specs of the P4? Things like 15 cycle multiplies, 1.5 cycle ADC/SBB, etc
And yet it's still faster than virtually any other processor on the planet. Intel has come a long way from when it was being spanked by MIPS/Sun/etc. You can make an argument for Alpha, but that's about it.
a non-x86 core
The P3, P4, and Athlon cores aren't x86. They have a wrapper layer that translates x86 instructions into their own internal core instructions, but that's it. And, frankly, a more "efficient" core doesn't make a bit of difference if it doesn't actually have any use in the real world. The x86 ISA is here to stay for a long, long time. People have been predicting it's death since it came about, and yet it's managed to dominate every other ISA out there. Hell, it's being looked at for embedded use now of all things.
Why should the CPU do the work of a compiler at runtime?
Because the compiler doesn't know what the dataset is. You can make guesses, but that's all. If you really want to optimize then you have to actually run the program for a period of time using real world data and then re-compile with the profiling data you've gathered. Which is pretty damned expensive to do, and is invalidated if your data set changes (yeah, that never happens in the real world) or you want to sell the program to multiple companies (again, one of those rare edge cases). The fact of the matter is that it's far, far cheaper to upgrade the hardware than it is to spend a bunch of additional programmer time optimizing the software. You can whine and kvetch about this, but it won't change reality.
Back when I was in college and was taking EE/CompE courses I couldn't believe how crappy the x86 ISA was either. And it is crappy. So what? It's still faster than everything else out there, it's cheaper than the competition, and the world has boatloads of software that runs on it. Do you have any idea how much software is used on a daily basis that hasn't been touched in years? How much do you think it would cost to replace all that software?
Don't worry. One day you'll graduate too and after a couple years in the real world you'll discover that a crappy solution that fits the job is far, far better than a perfect solution that doesn't do anything.
1. Place diamond wafers on pedestal. Depressurize chamber to one-tenth of an atmosphere.
2. Inject hydrogen, natural gas (CH4) into chamber. Heat with microwave beam. At 1,800 degrees Fahrenheit, electrons separate from nuclei, forming plasma.
3. Let it rain. Freed carbon precipitates out of plasma cloud and is deposited on wafer seeds.
4. Let it grow. Wafer seeds gradually become diamond minibricks, building up at half a millimeter a day.
5. Open chamber and remove diamond brick. Slice into wafers for semiconductors or cut and polish to make gems.
6. Profit!!!
DeBeers and Co. are very very unhappy about these two technologies and what they're going to do to diamond prices. Both companies can create perfect diamonds and the second manufacturing process will allow (once its been scaled up) for diamonds to be used in electronics.
But here's the reason the U.S. might just end up behind the technology curve:
Also, some other posters have commented on impurities being a stumbling block for diamond-based electronics, how convienent that "CVD diamond precipitates as nearly 100% pure"[Fuck Beta]
o0t!
The changes are largely necessary because of the unsavory consequences of Moore's Law, the famous dictum that states that the number of transistors on a chip doubles every two years
Moore's Law is only an empircal observation -- a convenient curve that fits through the our current data on time and transistor count. There are no gaurantees that this trend will hold for the future.
The point is that no physical phenomena forces the doubling. At best, one could say that mental and procedural limits prevent doubling faster than Moore's so-called Law. Perhaps this is the more interesting Law -- that doubling can't occur faster than every 18 to 24 months.
Two wrongs don't make a right, but three lefts do.
No, coppermine wasn't about using metal on gates. It was an all-aluminum (go figure!) interconnect scheme that used a low-k dielectric and thick wires for faster speeds. Only later did copper get used (and then again, only for the interconnects, not the gates)
HIV Crosses Species Barrier... into Muppets
Now how can you say that CPUs are based off of alien technology when Intel is making changes like this?
The same way we've always been saying it -- emphatically
--
The code name Coppermine had NO relationship with the metal used inside the chip. It was still an Al-on-Si chip, just like Katmai. Tualatin (last P-III core) and Northwood (second P4 core) were the first x86 Cu-on-Si chips from Intel (targeting Mobile/Server and Mainstream markets, respectively).
Additionally, AMD was making Cu-on-Si chips back at the Thunderbird (first "L2 cache on core" Athlon) debut. All cores that came from Fab 30 in Dresden were Cu-on-Si while all cores from Fab 25 in Autin were Al-on-Si. Palomino (first Athlon XP core) was made entirely at Fab 30 and thus all Palomino cores were Cu-on-Si.
IBM has been producing Cu-on-Si cores since 1998 (PowerPC 740, IIRC) and producing Cu-on-SOI cores since 1999 (PowerPC 750). Where do you think AMD got their SOI technology?
Moore's Law is a market imperative, which to a business is pretty much the same thing as a law.
Interesting insight. I wonder if there is an accidental collusion among semiconductor companies to limit their progress to Moore's observed trend? It seems suspicious to me that the trend should continue for so long without an obvious physical cause. In my orginal post, I suggested that mental and procedural limits kept companies for doubling faster than Moore's Law -- people just don't seem to create magic breakthroughs that double the transistor count in 3 months.
But now I wonder if Moore's law is a self-fulfilling prophecy. Everyone (semiconductor makers, software creators, and chip customers) knows about the Law, so everyone obeys it. Rather than spend time doubling the transistor count in a very short time, companies stick to the industry trend and spend time on other advances (e.g., innovations in microcode, cache, bus, branch-prediction, etc.)
The point is that in business, you need only beat your competitors by some incremental value. Thus, there is little incentive for Intel, for example, to double transistor count in 6 months as few customers would pay much more for the new breakthrough-density processor than they would for a competition-beating processor that only doubles on an 18-24 month schedule.
Perhaps Moores Law holds because everyone obeys it -- makers are too afraid to go slower and there's little competitive advantage to going much faster.
Two wrongs don't make a right, but three lefts do.