UIUC Creates World's Fastest Transistor Again

An anonymous reader writes "The University of Illinois has developed (again) the world's fastest transistor operating at over 500 GHz. They used an indium phosphide based wafer, and super-scaled dimensions. The device kind of looks like a spaceship." Milton Feng, the professor in charge of the team behind the transistor, admits that their ultimate goal is a terahertz transistor, which given their previous achievements, doesn't sound too lofty.

Comments

Cost break!

[Hegemony]

Sweet, now the 250 Ghz's will be totally affordable.

500GHz?!! I'll change my job!

[WARM3CH]

When I started designing hardware circuits, the world was much more beautiful. You could understand everything that your small micro-processor based system did, downto the function of the BJTs in the TTL devices down there... Then Intel started the 1GHz race and I had to learn a great deal of RF techniques to just design my next PCB. And now 500GHz?!!! At this rate, a few years later I'll have to learn more about RF and then eventually optics than next hot FSM synthesis algorithm! I guess I'd better change my job, start something more calm and steady, like paiting or ...

Improvement

[Carnildo]

From the article:
150 nm, 382 GHz
100 nm, 452 GHz
75 nm, 509 GHz

At their current rate of improvement, a 680GHz device will have a collector size of 0 nm. Just imagine what will happen once they manage negative sizes!

Re:Improvement

[spektr]

Just imagine what will happen once they manage negative sizes!

I imagine: 800i GHz in the first generation and even more imaginary in the following years!

Wise words from the chief developer:

[spektr]

The University of Illinois has developed again the world's fastest transistor operating at over 500 GHz

If only they had documented the damn thing, they wouldn't have to develop it twice!

Well, Duh!

[Anonymous Coward]

."The steady rise in the speed of bipolar transistors has relied largely on the vertical scaling of the epitaxial layer structure to reduce the carrier transit time," said Milton Feng, the Holonyak Professor of Electrical and Computer Engineering at Illinois, whose team has been working on high-speed compound semiconductor transistors since 1995. "However, this comes at the cost of increasing the base-collector capacitance. To compensate for this unwanted effect, we have employed lateral scaling of both the emitter and the collector."

I mean, that's just blindingly obvious.

Are you ready for lots of latency?

[G4from128k]

At 1 THz, it will take more than 40 clock cycles for a signal to move across a 1/2 inch die of the CPU. And it will take 320 clock cycles for a round-trip to a memory location just 2 inches away. (And that is assuming the signals travel at the speed of light in a vacuum, not the slower speed found in metal traces or optical fibers.) Should make it interesting for chip designers.

Re:Are you ready for lots of latency?

[RevRigel]

The speed they're talking about is typically GBP (gain bandwidth product), or the frequency at which the gain of the transistor is 1. It's not typically useful at a gain of 1 (for instance, if you want to fan it out to like transistors, it'll need to be at least n for n fanouts).
The clock speed on a chip is significantly slower than the speeds they're talking about because in order to achieve that external clock speed, the individual components must be faster. Say you had a P4, with its 20 stage pipeline. Each pipeline stage must complete in a clock cycle. However, say there's a propagation of say, 10 transistors for the output at the end of that pipeline stage to be valid. Each individual transistor would have to be 10 times as fast as the clock speed in order for the processor to work.
There will not be 500GHz or 1THz computers any time soon, at least not without extremely long pipelines and even faster transistors than this (to accomodate a useful fanout value).

Every time an article quoting a GBP-derived transistor speed comes out, everyone misunderstands this issue, so, here it is.

Misinterpreted

[Takahashi]

It seems like every time an article like this is on slash dot a million people say "wow I can't wait for a computer using that technology".

What people _don't_ understand is this is not the same technology as is used in a microprocessor. CPUs used Field Effect Transistors. The advantage of FETs is that there is no gate-drain current when the transistor isn't switching so they take very little power. With a bi-polar transistor, you are using a current switch, which would take massive amounts of current if you put many of these into an IC.

A more realistic application would be in communications systems where your carrier frequency is at 500Ghz.

Sorry to burst your bubble but you won't see 500Ghz computers next year. Maybe not ever using CMOS.

Re:RF is Obsolete?

[Bingo Foo]

So what's the vote: will RF designers be obsolete, or will digital designers have to become RF designers?

Ah, grasshopper: when you understand that the answer is "both" and "neither," then you will be on the path to entanglement.

Re:Improvement rate

[igny]

If you plot those 3 points on a plane you will see that the dependence is not linear. I tried to fit a curve through those points and got that

y=3000/x^0.4

where x is size (nm), y is speed (GHz). 1000GHz will be reached at ~15nm.

Re:How do you measure things that fast

[oobar]

You use the transistor to build a ring oscillator and measure the resulting frequency, then divide by the number of stages.