A Well-Chilled 750GHz Feasible Within 5 Years

drkhong writes: "...at least if you've got a good cooling system. IEEE Spectrum has an article about the next generation ICs. Using superconducting materials cooled down to 5K a peak of 750GHz has already been reached. Just think about how far light goes within one clock cycle, and then tell me you aren't impressed." These low-temperature devices are made of niobium (a superconducting metal), and use something called Josephson junction devices, resulting in chips for which the article states "there are no known physical barriers to decreasing size by a factor of 10 and thus increasing speed by a factor of 10, using lithography to move from today's 3-m linewidth to 0.3 m."

Comments

Re:open this box

[carleton]

Nope. That's the whole point of 128 bits. Assuming you're doing brute force, and have 1000 of these computers overclocked to run at 1000 GHz (to make the math easier), and of course assuming that they can do one trial per cycle,

2 ^ 128 trials * 1 cycle / trial * 1 second / (10^12 aggregate computer cycles) * 1 year / (3600 * 24 * 365) = 10790283070806014 years = 10 quadrillion years.

so what?

[crgrace]

Superconducting logic has been out for a VERY VERY long time. In fact, IBM burned tens of millions of dollars on the subject in the 1970s. The problems with superconductors are even WORSE than the problems with superfast III-V logic. UCSB has 70 GHz flip-flops made out of transferred-substrate heterojunction (III-V, Indium Phosphide and something else) transistors, but nobody thinks they will revolutionize computing, because they won't. So it is with superconducting logic.

There are two huge problems with superconducting logic that don't seem solvable in the near future. They are:

1. Cost : These things are enormously expensive to manufacture and operate, and it is the economy of scale of CMOS techology which has enabled, more than anything, the current computing revolution. Do you have any idea how expensive coolant and the dewar to use it in are to get something to 5K? Even the so-called "high-temperature" superconductors have to be pretty damn cold to function; they just don't need to go so close to absolute-zero.

2. Integration This is probably the killer. It will be extremely difficult to integrate many devices together. Even if myriad technical difficulties are overcome, the solution is not likely to be inexpensive, as CMOS technology is. For III-V semiconductors (which use much less exotic materials than superconductors), high defect rate, problems with lattic matching of the materials, and the lack of a high quality native oxide (like SiO2 in silicon) have made it impossible to achieve integration levels anywhere close to that achieved in silicon. Even GaAs, the most well-understood III-V semiconductor, can't be integrated to more than a few thousand devices. That's why we don't have 20 GHz GaAs microprosessors. And superconductors are even HARDER to deal with.

In summary, even if researchers are able to overcome almost insurrmountable odds to find away to reliably integrate meaningful numbers of these devices on a single die, I think it is very unlikely they will be able to do it cheaply, which is just as important as being able to do it at all. Otherwise, this technology will be of interest only to the military.

By the way, I know III-V semiconductors have a lot of very important uses, especially in optics and RF. It is a fact, however, that III-V logic is mainly of interest to the military and the space industry.

Re:750 Ghz

[krlynch]

I wonder if there are enough particles in the universe to run a finite elements simulation for more than 4 hours in a 750 GHz CPU.

There are more than enough to keep such a CPU busy for nearly all eternity....A 750GHz CPU (even assuming 1 flop/cycle average throughput) would still have a hugely difficult time just doing QCD calculations of the interactions inside a SINGLE proton in anything approaching days! (I have a colleague doing lattice QCD who was just telling me about their new algorithms for hacking time off of certain types of lattice simulations, and they are talking about running for 16 CPUyears on a brand new 90Gflop machine! At 750 Gflops, you're still talking 2 CPU years! Don't ask me for details, though, as I don't know any....not my field).

For further consideration, there are about 10^80 particles in the universe (give or take a few orders of magnitude.....). Let's assume it only takes 10flops to update a single particle for one timestep (not even close, but let's run with it shall we?) That means we update 75 x 10^9 particles every second...let's round up and call it 10^11. That means it would take about 10^69 seconds to update one time step. Or 10^61 years. Which is roughly 10^45 times the age of the universe. Not to mention the amount of RAM you'd need to run this simulation on (which would take more particles to build than there are in the universe itself, but I digress.....)

Really monstrously fantastically mind-bogglingly large numbers are really really fun :-)

Cheaper Hardware?

[dmatos]

With the advent of the 300MHz processor, the 233 I purchased became dirt cheap. Now that there are 1.5 GHz chips out, you can get an 800 MHz chip dirt cheap. When the 750GHz chips are produced, I will be lined up to buy an obsolete 500GHz chip that will be fast enough to start windows from boot in less than three minutes! Yay bleeding edge subsidizing second-stringers!

Yeah, and a 2 million stage pipeline

[Mr Z]

One problem with these high clock rates is that you end up having to pipeline things rather excessively all over the place. I'd imagine at 750GHz that even a single 64-bit ADD would be pipelined over multiple cycles, due to transport delay!

Think about it: Light travels about 1 foot per nanosecond (30cm). At 1GHz speeds, a signal could travel well across a die if it were unimpeded (eg. could travel at the speed of light). In fact, it could theoretically travel most of the way across the motherboard in one clock period. At 750GHz, light travels 0.4mm per clock tick -- about 1/20th the way across a typical CPU die (assuming a die in the range 8mm x 8mm to 10mm x 10mm die -- not too far off what we build today). We're talking 20 pipeline stages just to get from one edge of the die to the other, if we can travel at the full speed of light in a vacuum. And the bad news is that we probably can't -- just look at todays CPUs!

What'll happen is that highly parallelizable problems will speed up, and inherently serial problems will end up staying the same. All of your number crunching for playing video games will rocket along since the calculations can be pipelined and parallelized, but the twisty, turny, five-instructions-and-a-branch control code won't speed up much.

--Joe
--
Program Intellivision!

WOW!

[clinko]

Joe Consumer -

"750GHZ! WOW! NOW I CAN RUN AOL EVEN FASTER!
AND WITH 56k AOL IS FASTER THAN EVER!"

Finally!

[Reality Master 101]

I can get 100,000 frames / second on Q3. Dammit, I can see the difference!!

--