Slashdot Mirror
Billions of Transistors on a Single Chip
cgi-bin writes, "IBM has reportedly developed technology to create "tens of billions" of transistors on a single chip. Intel's pentiums only have 27 million or so. The technique uses electron beams instead of the traditional optical lithography. "
What's interesting is how some of IBM's current processors (such as the PowerPC 604e, 750 "G3", and once Motorola lets em, the 7400 "G4") really don't have that big of a transistor count compared to what Intel and AMD have. Especially when you compare performance. Imagine what they could do...
Second, and perhaps more importantly, the same techniques used for the etching -could- be used to produce ultra-high definition TV. (After all, all you're using is a somewhat larger electron gun.)
Now, I don't know about you, but I like the sound of computer monitors (or domestic TVs) capable of definitions of up to 500 billion lines. So what if nobody would be capable of telling the difference - it's the coolness that counts, not the practicality! :)
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
The advantage here is that realistic amounts of memory--128MB or more--can be put on chip with the processor. In effect, all memory is cache. This would be fantastic both in terms of speed and low cost.
If you can make processors with this technique, I'm sure you can also make memory with it. So you would still have off chip memory that is far bigger than cache. So instead of 512k cache and 128M main memory, wouldn't you have 128M cache and a few gigs main memory...
--
bgphints - internet routing news, hints and ti
Of course, with feature sizes this small, processor speeds would be improved, too. So both interpretations are valid. But I believe they indeed said what they meant.
I can see the fnords!
When used in Fields and Waves (EE313 iirc) then E, as in "the loop integral of E dl equals 0," is often written in script or boldface, and represents Electric Field strength, in Volts/meter.
And E as in E=MC^2 has already been covered in a prior reply. I think most applications are in Megajoules, though.
Anyway, none of those really apply to the original comment, which was concerned about heat. The applicable law here is named after, erm... Watt, I believe: P=IE. When it comes to resistive heating, applying ohm's law lets us express that as P = I^2 * R. Deriving the units for P is left as an exercise for the student.
I can see the fnords!
Given the time it takes to build and configure and initialize a fab plant, this is actually a rather short period of time. It takes anywhere from 4-7 years to gt a new fab plant fully operational, with 5 years the norm. Replacing the fabrication hardware, still takes many many months to years to get installed and setup.
How do you know our conciousness isn't 'embeded' in our genetic make-up
Well, it is certainly influenced by it, but if that were the case, identical twins would have an identical consciousness, which we know not to be true.
However, identical twins are often very similiar in behavior, even when separated at birth and raised in different environments. So genetic predisposition definately has a part in the development of your consciousness, personality, and intelligence.
Dont forget that a refined implementation of this technique would yeild lithographic etchic at the near atomic level .
MMmmmmm nano nano
Your thinking too small!
With this type of size reduction we could have chips at 3 gigahertz with 50 GIGABYTES of L1 cache. At bootup your system would load your entire drive image into L1 memory, and there would never be page faults or disk hits. Except to save, but that would be a background process and not affecting computing.
Now THAT would be a fast system.
Sir, what you need is a cubic meter crystal of computational ecstasy.
"All tasks can be designed / computed / evaluated / indexed / summarized / optimized / etc by a solid cubic meter crystal super conducting FPGA like nano scale self assembled ceramic block running at 90 gigahertz and chilled to 20 degrees kelvin running evolutionary systems of simulations / knowledge agents / neural networks / genetic algorithms / bayesian probability maps / self organizing pattern matching systems. All types of evolutionary systems layered and nested to unlimited levels of abstraction and complexity. Hardware capable of performing a centillion teraflops and holding a centillion terabytes of memory in core, operating at crystal speed. "
yes, that is what im waiting for...
Well, not really.
You could squeeze the instructions for building a physical body like yours onto a CDR, but your body is uniquely yours, and to save its configuration (meaning the output from the genetic program at this point) would still require orders of magnitude more information. Sorry, i misread your comment, i was under the impression you were saying you and your mind..
True. I wish I could find out exactly how fast this process is going to be. If it is going to take three days to etch my billion transistor wafer, then Id rather suck a donkey.
Well, you could squeeze your entire physical blueprint onto a CDR. However, the thing that most of us value most, namely our consciousness, would require many orders of magnitude more space.
If you had read the article you would know that this technology would be quite inexpensive (for a fab budget) to implement with practical applications immediately.
This isnt far out theoretical research, they have built the system, and scaling it to fab production sounds quite straightforward and practical.
That was UNTIL they figured out a different way to do it.
You forget the fact that this type of production is to be much faster than current optical methods. That would be where the real big payoff would be. Higher chip density is a plus, but coupled with fast fab would be great news for the industry.
Funny how Carl Sagan is remembered for that phrase -- when it was actually part of an act that Johnny Carson (?) came up with to make fun of Sagan. I gather hearing the phrase -- especially hearing it falsely attributed to him -- got really tedious to Sagan, though he eventually learned to laugh it off, and even used it to title his last(?) book.
Electron beam lithography is nothing new, nor is IBM the only one developing it. In fact, there have been much smaller transistors made (such as the 18 nanometer transistor made here at UC-Berkeley using e-beam lithography).
The drawback to direct-write electron beam lithography is that you have to directly trace the circuit you are trying to print in most cases, while in optical lithography you can expose an entire die (or multiple die) at once. There have been improvements made over the years, using techniques such as parallel writing, but it's still slow. Even using a more conventional masked resist and scanning the beam across the wafer using vector or raster methods, there are problems with electron scattering and such.
This article is pretty short on technical content, so it may be that IBM has developed a way to make e-beam lithography fast enough to be used in a production environment for chips (it is already used for making photomasks). That would definitely be a significant development. We'll have to wait and see, I guess.
Also, keep in mind that just because they have a lithography tool that can write 80 nanometer lines does not mean that the rest of the processing equipment (etching, planarization, etc.) could support it. There would need to be advances in those tools as well.
My other question is, what do we do with tens of billions of transistors? If we jump three orders of magnitude in the number of transistors on a chip, is it really going to do us any good, at least with current circuit design techniques? I think testing a circuit like that would be a nightmare.
-Jason
Please alert me if I am wrong, but IIRC the smaller the transistor, the lower the power requirement (less heat), and the faster the chip (less distance from junction to junction). So if all they did was to make the same chips we have now on the smaller die size, there would be a reduction in power requirements and a speed increase, right?
Not that I'm much of an expert in these things, but when IBM says they've got the tech Nikon has built and demonstrated a a proof of concept machine, this sounds like tech that's less futuristic than say, quantum gates, etc. I mean, Nikon isn't in this to produce a one-off demo machine -- what they're really after is the ability to put their machines into the fab plants. So the actual production of chips is probably still a couple of years off, but the technology would vault IBM ahead of just about every other chip maker on the planet -- ahead of Intel, Motorola, AMD, TI, and anybody else I may have forgot.
My biggest remaining questioss not answered by the article are:
- with the smaller wires, is there a higher crosstalk problem at higher switching speeds, and
- Once they've got the tech completed, will it be an IBM only tech, or will it be something they license so that the rest of the world can benefit (Personnally, I'd like a faster Transmeta chip, a faster StrongArm, etc.)
Let's hope this tech pays off and we all see the benefit soon....Open Source isn't the only answer -- but it's almost always a better value than the alternatives...
The father of nanotechnology spoke in 1959 of etching with reversed electron microscope. He also refers to us in the year 2000 looking back...
Funny how I remember Carl using this phrase repeatedly in the PBS series "Cosmos", but I never remembering hearing it from Johnny Carson (of course, I never really watched Johnny Carson).
Aah, change is good. -- Rafiki
Yeah, but it ain't easy. -- Simba
I came to this point one night. Why aren't we using light? Light is faster! Eventually we will be limited to the speed of electricity.
Then I realized that you can't escape electricity. You would have to convert the light beams with optical switching sooner or later in the computers, and that would be an insane performance decrease.
You cannot escape the fact that memory will always run on electricity. It just isn't practical to come up with a type of memory that is based on storing light.
Also you run into the problems with refraction and the fact that if you want to make sure that each light beam doesn't interfere with the other beams during a processor cycle you would have to make the processor very large so you could sheild the indivudual light beams with an opaque material. Also, lasers get very hot. Its just not practical.
The advantage here is that realistic amounts of memory--128MB or more--can be put on chip with the processor. In effect, all memory is cache. This would be fantastic both in terms of speed and low cost.
The wavelength of visible light is already larger than the feature size of current generation electronics, so if we were to move over to opto-electrical computers they would actually be a lot larger (and thus slower because of the time taken for signals to travel anywhere). However using small lasers to communicate between chips could have some potential.
Also as the other reply states, any useful optical device actually involves electrons transitions anyway to provide some kind of nonlinearity (you can't build anything useful such as a NAND gate out of linear componants).
So has anybody found a URL which explains what advances IBM have actually made which makes this better than the E-Beam lithography that's been around for ages?
Edmund Green.
Nanoscale Physics Research Laboratory, The University of Birmingham, U.K.
>It just isn't practical to come up with a type >of memory that is based on storing light.
No, it is practical to come up with the idea. It's just not practical building and using it effectively.
Researchers at University of Colorado, Boulder designed an optical computer several years ago. There are several systems in existence where computation is partially or fully done optically, but this was the first (and the only, if memory serves me right) system to do everything optically-i.e. it had a memory system based on storing values optically.
The system essentially stored pulses in a loop of fiber optic cable several miles long. I think the principle is analogous to very early electronic memory systems were bits were stored in forms of waves on tanks of mercury.
Extensive info on this was published in some IEEE publications back then. I don't have the time to look for the URL now, but it will be helpful if someone can find the reference to it.
--
BluetoothCentral.com
A site for everything Bluetooth. Coming soon.
Zigbee Central: A Zigbee weblog
Here is the link to the optical computer that I mentioned. The paper has a description of the optical memory system and logic structures.
Stored Program Optical Computer(SPOC)
--
BluetoothCentral.com
A site for everything Bluetooth. Coming soon.
Zigbee Central: A Zigbee weblog
Called PREVAIL, ... the technology would ... significantly improve the speed at which silicon chips can be processed, researchers said.
I bet What the researchers probably said was "significantly improve processor speed." This is an important point to make because anyone in the semiconductor industry knows that electron lithography is SLOW, like orders of magnitude slower than optical lithography. That is why nobody has ever used it to make commercial chips even though the technology has been around for more than a decade. I would be interested to see some more technical back-up articale that talk about masking and throughput.
"s another side note to this, Lucent Tech. has an EBL system just about at proof of concept called SCAPEL. Hope this clears up a few of the wrong ideas and helps people understand what this is all about."
Might I point out that here on slashdot back in October or November was an announcement that Lucent had reached a resolution of .05 microns using SCALPEL. Then let's not forget the UC Berkeley student who made an even smaller transistor two weeks later. Something like 0.018 microns or so. This IBM announcement is not really anything new. Here is the press release.
With this technique you could build a do-anything chip based on Transmeta's technology which has dozens of cpu cores and an extremely powerful graphics processor. And the chip would still be tiny, ridiculously fast (many gigahertz), cheap, and still under 1 Watt. I need one!
Cool! I could just squeeze my entire (physical) self onto a CDR!
---
DO NOT DISTURB THE SE
Actually, this is not the only technology that promisses to bring lithography to even lower scales... There's also Lucent's Scalpel, Intel's Extreme Ultraviolet (EUV) and an X-Ray technology from IBM. There all saying they're stuff is better that all other technologies... let's see which on comes first
Opus: the Swiss army knife of audio codec
It's all great and wonderful if they can construct devices that small, but the question is whether they will even work or not. Granted, the theoretical minimum device channel size is far below what is currently being produced (I think it's somewhere in the range of 0.02-0.05 microns, but don't quote me on that).
Another problem to look at is the degradation of the device that can take place when things get that small. I'm sure people wouldn't be so prone to overclocking their processors if there was a chance they might completely destroy the processor in doing so.
Another key, as was already stated, is that they need to bring the cost of the process down before it will ever see a production line. If the process requires a long time then it's likely to either create a bottleneck in the production line, reducing the overall output, or simply drive the price of the final product up a bit by forcing the company to purchase large numbers of the tool that performs the process.
Granted, the savings that would result from a smaller die size and potentially a correspondingly small package size could make up for the price difference due to the new tools. I'm not sure of the exact number, but a large part (>50% I believe) of the cost of the chip is in the packaging (Which is why you'll find bins of scrapped wafers at any production plant.. why package something that isn't going to work)
Another problem I could see in bringing the process to market is in contamination of the chips during production. As it stands now, lots of chips are scrapped because of skin cells, dust, etc landing on them during their trip down the line. With the smaller device size the smallest foreign particle size that could be tolerated would have to be smaller... so either clean rooms would have to get cleaner and their employees more religious in following the rules, or they would have to find some way to isolate the wafers from the technicians.
This sounds to me like it's electron beam lithography, but not SCANNING electron beam lithography.
Electric fields can be used as lenses to focus electron beams, forming images of a stencil, just as physical lenses can be used to focus photon beams.
On one hand there's a complication because electrons mutually-repell and also affect the field that forms the lens, so higher beam currents tend to distort things somewhat.
On the other hand, the lenses are formed by an electric field's natural curvature. So small-scale optical imperfections just don't occur in a good vacuum, while gross imperfections are easy dealt with by maintaining decent tolerances in the construction and excitation of the electrodes.
Of course they COULD have made a breakthrough in scanning electron beam technology, and be talking about writing every chip one at a time. But that doesn't square with either the claims of "billions of transistors" and those of "speeding up the processing".
Yes, they could get DENSITIES of billions of transistors. But writing them one at a time takes a while. And keeping the beam alligned across a large chip is a problem. (Though the latter can be solved to some extent by first laying out a set of location markers and using them in later steps to figure out where the beam actually is.)
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
I can hear Carl now from his grave.... "And this chip is populated with billions and billions of transistors on a single chip, all interacting and dancing to a tune called ... COMPUTING!"
---
DO NOT DISTURB THE SE
John Markoff has a somewhat more detailed article on PREVAIL here.
The good news:
The new chip has over ten billion transisters.
The bad news:
The new chip is over 1700 square feet.
Plans for a portable based on the new chip are being put on hold...
The cake is a pie
1. Inorder to get the resolutions required in the future, photon lithography would have to go to X-rays that have a high enough brightness (i.e. You need a syncrotron X-ray source on site). For EBL, you just need a large source filament.
2. Masks for X-rays would need to be the same size as the actual features because there is still not a good method to make images out of X-rays so they use a shadowing technique. Not the case for EBL. Electrons use magnetic lenses to focus and have been used and designed for years. This actually allows you to build the mask in seperate parts and have the electron beam deflection put it all together for you as if it was all together.
3. Stepper motors don't need to be quite as acurate on positioning. This is because you can put in a simple feedback unit that examines where you are projecting on the surface of the wafer and deflection coils can position the beam exactly. This means that you can do lithography while the wafer is still moving! You couldn't do this in your wildest dream with X-rays.
4. Electrons have a very small wave-length at the acceleration voltages used (on the order of picometers). However, the real limitation for EBL is not the wavelength by lens abberations (pick up a good optics book) as well as space-charge effects (this happens because you are using a charged particle and they repel each-other giving a bluring effect). Even with all of this, some predict that you could get resolutions "easily" to the 10nm scale in lithography. No, we can't do atom manipulation with this technique.
5. No this technique does not use a focused beam technique (similar to scanning transmission electron microscopy), but it uses a plane-wave electron beam so that you can expose large areas at once (similar to stardard transmission electron microscopy), allowing for higher through-put.
Probably the major disadvantage for EBL right now is that we need more sensitive resists. The brightness of the EBL is still low compared to UV photon lithography, but I know of several groups that have come a long way with this one.
As another side note to this, Lucent Tech. has an EBL system just about at proof of concept called SCAPEL. Hope this clears up a few of the wrong ideas and helps people understand what this is all about.