Monday, February 08, 2010

That Is Fast


Carbon (graphene) transistors are getting really fast.
IBM Research has demonstrated a 100GHz transistor. Fabricated on new 2-inch graphene wafers and operating at room temperature, the RF graphene transistors are said to beat the speeds of all but the fastest GaAs transistors, paving the way to commercialization of high-speed, carbon-based electronics.

"There are all kinds of extraordinary claims being made every day for graphene semiconductors, but this is the first demonstration of a RF graphene transistor that was made under technologically relevant conditions and scale," said IBM Fellow Phaedon Avouris, who oversees carbon-based materials efforts at IBM Research.

The graphene RF transistors were created for the Defense Advanced Research Project Agency under its Carbon Electronics for RF Applications (CERA) program. Almost four times faster than previous demonstrations, the graphene transistors were fabricated at the wafer scale using epitaxially grown graphene processing techniques that are compatible with those used to fabricate silicon transistors.
Big transistors are are not too hard. You essentially lay down a sheet of graphene, dope it (or dope it while it is part of the SiC substrate), and then put a gate pattern over it. You get a power transistor. At 100 GHZ that probaly is indicative of the ability to work at 50 GHz.

But they have set their sights on bigger game.
There are several relatively easy steps to further widen the gap between graphene and silicon. For instance, graphene suspended over an air gap and supercooled has achieve carrier mobilities of up to 200,000cm²/Vs compared to silicon's 1400cm²/Vs.

IBM's demonstration of room-temperature graphene on an insulating substrate only achieved 1500cm²/Vs.

The gate length of IBM's graphene transistor was 240nm, nearly 10x larger than the smallest gate lengths achievable with current lithographic techniques (under 35nm). By optimizing its process to increase mobility and shortening the gate length, IBM will next aim to increase the speed of its graphene transistor up to 1THz, which is the goal for the CERA program.
Computers are a little different. Take the top speed and divide it by 4. Then allow for 6 levels of logic (AND, OR, and NOT) plus wiring delays and you can divide that number again by 10. So 1 THz/40 = 25 GHz. About 10X faster than today's computers. If they can cut the heat load by a factor of 5 to 10 they will have one screaming machine. Who will be the early adopters? Gamers, server farms, and of course the guys who funded it. The military through the CERA Program.

4 comments:

Anonymous said...

Is the heat load solution just more of what is currently used - fans, circulating coolants etc.?

I have often wondered if there is a way (currently or in principle) to use high powered computing to reduce the amount of data that needs to be sent when transmitting data. From game lag to long download times, broadband is never really broadband as new applications use it up. If the power of computers at either end of a line (which are probably easier to upgrade than installing improved cable) could somehow process the info into smaller amounts... I know that some of this is already done like data compression protocols. I guess what I am curious about is to what degree data compression can be advanced by raw computing power increases.

M. Simon said...

Bandwidth is cheap. My pipe (at no extra cost) give me US continental unlimited calling at no extra charge.

I used to get 180 minutes a month with my plan.

There is not a lot to be gained from data compression. Video is already compressed by one of the mpeg standards.

Susan's Husband said...

ByteMobile would disagree with your assessment of the benefits of compression.

As for server farms, why would they care much about faster processors? Bandwidth in its various manifestations is their limiting resource. Multi-core processors is what helps them out.

With regard to heating, you can compute a "MIPS per watt" value for various processors. If graphene has a higher base value, then you get more processing with the same amount of cooling. I suspect this is what M. Simon meant by graphene ameliorating the heat load problem.

M. Simon said...

If this is the ByteMobile you were thinking of then it is more a reconfiguration of the compression rather than better compression.

Remember:
Audio: uLaw or ALaw.
Video: mpeg
Text: does any one actually compress it?

What I'm saying is that with most (all?)streams precompressed you can't save many bits by another compression scheme.

The ByteMobile guys take advantage of the users. Which is different from better compression. They send the bits when they are needed not at first demand. There are server suites that are working this angle too.