Slashdot Mirror
Graphene Won't Replace Silicon In CPUs, Says IBM
arcticstoat writes "IBM has revealed that graphene can't fully replace silicon inside CPUs, as a graphene transistor can't actually be completely switched off. In an interview, Yu-Ming Lin from IBM Research (Nanometer Scale Science and Technology) explained that 'graphene as it is will not replace the role of silicon in the digital computing regime.' Last year, IBM demonstrated a graphene transistor running at 100GHz, while researchers at UCLA produced a graphene transistor with a cut-off frequency of 300GHz, prompting predictions of silicon marching towards its demise, making way for a graphene-based future with 1THz CPUs. However, Lin says, 'there is an important distinction between the graphene transistors that we demonstrated and the transistors used in a CPU. Unlike silicon, graphene does not have an energy gap, and therefore, graphene cannot be "switched off," resulting in a small on/off ratio.' That said, Lin also pointed out that graphene 'may complement silicon in the form of a hybrid circuit to enrich the functionality of computer chips.' He gives the example of RF circuits, which aren't dependent on a large on/off ratio."
Can't be switched off? We finally found perpetual motion. It's huge!
because I'm using a silicon-based CPU
Silly people, you can't make AI with silicon. Everything is 1 or 0, on or off, yes or no, true and fals....
With the introduction of graphene we can have "Maybe" and "False but I'm might be lying" or "No thanks this is boring"
So why not use it in a current-switching logic like ECL? Yeah, the power consumption might be a bit high, but you only need to make the cache and cores run faster.
That is all.
The thing that strikes me about this is that when something new comes out that has different strengths and limitations, at first we are at a loss for how to make it useful, but then we work around the limitations and are much better off.
The things I'm aware of that make the contrast between high and low important is measurement of those highs and lows and short term memory. So those are potential areas for improvement that could make the technology viable.
Just because they haven't worked out how to do it yet, doesn't mean it can't be done.
Funnyhacks - Wierd, unusual, and fun hacks
100GHz? 300GHz? 1-fucking-THz?
I started drooling just a bit. Talk about a jump in speed. That's like going from floppies to thumb drives. Maybe even portable hard drives.
Sent from my CR-48
So it can't be "completely" switched off? What does that mean even? That almost sounds like the perfect application for SSD storage.
Life is not for the lazy.
Why they're right:
Graphene is a metal (or semimetal, whatever). Capacitive effects cause your current gain (ratio between input current and output current) to drop with frequency. The highest practical frequency of operation is where the gain is 1. IBM demonstrated a year ago a transistor whose current gain reached 1 at 100GHz (also known as fmax or unity gain frequency). However, that's just current gain. Digital circuits require a voltage gain as well. You can have high current gain but not have a high voltage gain by having a low output resistance.
Why they could be wrong
Without giving a crash course in electronics, computer (not all) transistors these days operate by raising/lowering a potential that either "gates" electrons and keeps them from passing through or allows them to go through freely. How big the gate is depends upon something called the band gap size, i.e, the energy required to move an electron from an atom (valence band) into participating in macro-scale conduction (conduction band). Graphene does not have a band gap.
However, you can artificially create a band gap by excluding the low energy, long wavelength electron states that exist at the bottom of the conduction and top of the valence bands. You do this by patterning your graphene into strips below about 30nm in width. In this way, no electron state with a wavelength longer than the width of the strip can exist. In a few years (right when graphene starts to hit its stride outside academia), such patterning will be possible (Intel is at 32nm now, if you recall).
As IBM breaks out the bad news, causing chip R&D departments of competitors to halt research into graphene, IBM releases a new 500GHz processor next year.
from 09 F9 11 02 9D 74 E3 5B D8 41 56 C5 63 56 88 C0
to 45 2F 6E 40 3C DF 10 71 4E 41 DF AA 25 7D 31 3F
and I was looking forward to drawing my own cpu's
Diamond, on the other hand, has a band gap of 5.5 eV, and has _excellent_ thermal properties.
Put my fist through my alarm clock with its ding-dong death inside my ear. - The Blackjacks.
Most of the freaking chip is cache. Have a look at the floorplan sometime.
Intel engineers sometimes joke that they're the biggest memory vendor that nobody heard of.
The fundamental problem with "doesn't turn off" is that leakage current (IDS(OFF)) is already a major component of chip dissipation, even when we use all sorts of tricks to reduce it. With graphene, that goes from "problem" to "useless." Except for analog, where the transistors don't turn off anyway.
Lacking <sarcasm> tags,
We have been using hafnium oxide instead of silicone to make chip for several years now. Silicone can't handle the tight lattice distances at such small scales anymore.
I called it a mighty Sperm Whale, she called it Finding Nemo.
Someone tell this guy about "noise" and "filters" and the like. On or off. That's a good one, lol.
All right. A digital signal can be thought of as an analog signal plus a "quantization noise" that's the equivalent of the difference between the digital and analog signals.
When you filter out the noise from the digital signal you have the analog signal back. The interesting part is that it you can shape the noise so that it can be perfectly filtered out of the digital signal.
In other words, that old "wisdom" that analog signals have infinite precision, while digital signals are limited by their resolution is not so true as it may seem. With correct digital signal processing, a digital signal can be just as exact as the equivalent analog signal.
because after I seen the statement on gap, I searched and immediately got an article - graphene not only can be equipped with gap, but tunable one. just make double layer graphene: http://spectrum.ieee.org/semiconductors/materials/graphene-makes-transistors-tunable
This is covered somewhat in this spectrum article:
http://spectrum.ieee.org/semiconductors/materials/graphene-electronics-unzipped/0
Too bad we can't fashion graphene into a full replacement for silicon, but only a complement. Unlike silicon, graphene is ambipolar, which means it lets current go through regardless of whether a positive or negative gate voltage is applied. Doped silicon, by contrast, is unidirectional—that is, it allows current flow only when the gate voltage is at a certain polarity (say, positive or negative). Ambipolarity may seem like a bug rather than a feature, but with a little thought, electrical engineers can design around this problem and even create devices that take advantage of it. Researchers at MIT have demonstrated an elegant example of this by using ambipolar graphene FETs as frequency multipliers, which double the frequency of an electromagnetic signal in radio communications and other applications. (This device employs pure graphene FETs, enabling it to perform a function that carbon can handle but silicon cannot.)
The other problems with graphene mentioned aside, you start to run in to speed of light issues with extremely high frequencies. At high frequencies the wavelength is so short, that it can't travel across a chip in a single cycle. That has some real design issues. For example at the full speed of light a 5GHz signal has a wavelength of 6cm. Ok, not a problem. A core in a CPU is smaller than that so the signal can travel anywhere in a single clock, even taking in to account that wire runs could be longer. However at 50GHz, well then you are only talking 6mm. That's a potential problem. Current chips are larger than that, never mind the wire runs. Maybe if cores are kept small and simple it is fine, but it is getting problematic. At 1THz you are talking only 300 micrometers wavelength.
So even if graphene becomes practical, speeds that high may never make it in to CPUs. That a transistor can operate at those speeds doesn't mean a whole CPU can.
Could analog computing solve this problem? If it can't generate a digital signal, then could it generate something that an analog computer could interpret, possibly to much greater effect than in a digital system?
My understanding is that analog computing is gaining ground in the research field (after being dormant for decades) and this seems like perfect timing.
while(1) attack(People.Sandy);
>Graphene is a metal (or semimetal, whatever). Capacitive effects cause your current gain (ratio between input current and output current) to drop with frequency.
A metal has a clear overlap between conduction and valence band. Graphene has zero band gap only for k=0, putting it right in between a semiconductor and a metal - hence semimetal. It clearly is not a metal.
The capacitive effects are largely unrelated to the bulk properties, but to the gate insulator of these devices and various parasistics. There is something called diffusion capacitance, which scales with electron mobility. The mobility in graphene is unusually high, allowing for a faster drain current control.
> However, that's just current gain. Digital circuits require a voltage gain as well. You can have high current gain but not have a high voltage gain by having a low output resistance.
Without giving an in depth explanation of how CMOS circuits work - I am not sure you are very familiar with the specifics?
>How big the gate is depends upon something called the band gap size, i.e, the energy required to move an electron from an atom (valence band) into participating in macro-scale conduction (conduction band). Graphene does not have a band gap.
This is pure bullshit. I really don't know how to put it. You are relating two things which have no relation. It is like saying "bananas are yellow because apples grow on trees".
>In a few years (right when graphene starts to hit its stride outside academia), such patterning will be possible (Intel is at 32nm now, if you recall).
There is something called band-to-band tunneling which allows electrons to transition from conduction band to valence band even if there is a gap and a field exists within the semiconductor. If the band gap is high enough, and preferrably indirected, this effect does only play a minor role. For low band gap materials, this effect would dominate junction leakage and would effectively prevent them from turning off.
This is wrong and right - a single layer of graphene has no gap and one can only be produced if it lies on an appropriate substrate (one that introduces an energy difference between the two different basis point within the graphene lattice). However bilayer graphene exerts a tunable band gap with the application of an electric field - essentially the conductive state is gate controlled, the problem at the moment is the on/off ratio which is being hampered by a number of things including the cleanness of the graphene and the size of the electric field you can apply. I can assure you it's very possible to make transistors from graphene - I've done it.
"Oh boy"
Wow... A lot of this is missing the main issue that has now limited the speed of processors for a number of years - Limitations of the interconnect.
The average microprocessor is not limited by the capability of the transistor, but rather the RC time constants associated with the connections between them.
Thats the reason, aluminum as a metal interconnect was dropped a while back in favor of copper. Lower R for the same C.
Analog computing? good luck with that!
www.effectiveelectrons.com "chips that work" Analog, RF, Mixed Signal