TeraHertz Chip

[DarfBomb]

Me and my friend were talking the other day and we got onto the subject. He said that the government was working on (or had, i can't remember) some kind of supercomputer processor that ran in the terahertz range. I'm not sure where he herd this but I'm curious so my question is if a chip like this is even possible. If so what kind of archetecture/tricks of the trade do you think would achieve this speed and what would this chip be used for.

 

[CTho9305]

You probably mean "teraflops", not "terahertz". A cycle time of 1 picosecond isn't remotely feasible even on the most advanced manufacturing processes today (a single gate takes more than 1ps to switch). Teraflop computers, however, have been around for a while - check out top500.org for a list of the fastest supercomputers. Note that these supercomputers are usually built from hundreds of small machines which do things in parallel. If you don't have a task that can parallelize easily, you won't get good performance from them.

 

[f95toli]

I know of one such project. It was known as the "teraflop"-projekt and was based on superconducting RSFQ-technology. The research was done at Stony Brook (SUNY). Head of research was Konstantin Likharev. AFAIK they only managed to make a few small-scale demo-chips before the project was closed down in 2001 (or 2002, I don't remember), they had a lot of problems with their fabrication (reproducibility).

The american miltiary closed their foundry for RSFQ-technology about a year ago and now they use civilian foundries instead. However, RSFQ is still a very active field of research and companies like Northrop-Grumman are still doing R&D, presumably for miltary applications.

Edit: I should point out that the goal of the teraflop-project was to make a single-chip CPU, it was NOT based on parallell processing. I.E. the clock-frequency would have been 1 THz.

 

[AluminumStudios]

If you read up on the challenges that Intel and AMD face in bumping up the clock speed the answer seems to be no - you can't have a 1 THz chip. The laws are physics are against you. The time it takes electrons to move from place to another plus the times it takes gates to open and close plus the capacinance, resistence, and current leakage issues would be pretty much insurmountable with electricity based chips.

Most companies have realized that physics is kicking in and computers rather than getting faster have to get wider and do more in parallel - hence the multi-core Pentiums, AMD chips, IBM Power chips, and other's like the CPUs of Cray's X1 systems.

It's concievable that an individual chip could hit 1 terraflop of performance though. In Anandtech's current article on Xbox360 they mention that MS is claiming that it's CPU can do 1 tflop.

 

[f95toli]

Intel and AMD are still using silicon, it is relatively straightforward to increase the clock-frequency if you use other materials. III-V semconductors such as GaAs, GaN and InP are much faster than silicon.

In the teraflop-project I mentioned the plan was to use niobium which is a superconductor, NB-based RSFQ circuits can be clocked at several hundred GHz. However, fabrication is still a problem and it is difficult to make chips with more than a few thousand elements.

 

[Chode Messiah]

U know, some guy took like 50 itx mobos and ran 'em in parallel in an atx case. He said he got over a teraflop out of it. (i think the cpus were P4s)

 

[imported_Phil]

Originally posted by: Chode Messiah
U know, some guy took like 50 itx mobos and ran 'em in parallel in an atx case. He said he got over a teraflop out of it. (i think the cpus were P4s)

Link please, because not only would you not be able to get fifty ITX boards in a regular ATX case, this link says that they've only ever used VIA EPIA/C3 chips, not P4s.

 

[CTho9305]

A 600MHz Athlon can do 2.4gflops, so you'd need 417 of them to get a teraflop. Of course, since you can get 2.4GHz Athlons, 105 would suffice, but that's the absolute peak theoretical performance. You'll never see it in the real world. The easy way to get a teraflop (though not general purpose) would be to put together 5 boxes with top-of-the-line video cards (which can do >200gflops each).

 

[AluminumStudios]

Originally posted by: CTho9305
A 600MHz Athlon can do 2.4gflops, so you'd need 417 of them to get a teraflop. Of course, since you can get 2.4GHz Athlons, 105 would suffice, but that's the absolute peak theoretical performance. You'll never see it in the real world. The easy way to get a teraflop (though not general purpose) would be to put together 5 boxes with top-of-the-line video cards (which can do >200gflops each).

I'm afraid parallel processing doesn't scale that well or work as simply as you've stated. If you go look at #401 on the Top500 computer list (http://www.top500.org) you'll see it's just shy of a terraflop with 900 2.0 GHz Opteron processors.

Memory bandwidth, inter-cpu communication, and other latencies really severly cut performance as you scale upswards in CPUs. Actual performance is MUCH LESS than the sum of what the CPUs are capable of.

 

[Calin]

Originally posted by: Phil

Originally posted by: Chode Messiah
U know, some guy took like 50 itx mobos and ran 'em in parallel in an atx case. He said he got over a teraflop out of it. (i think the cpus were P4s)

Link please, because not only would you not be able to get fifty ITX boards in a regular ATX case, this link says that they've only ever used VIA EPIA/C3 chips, not P4s.

Also, VIA EPIA are socket-compatible with Pentium3 processors, not Pentium4. And I remember the article, he had a rack with mini ATX mainboards, not a computer case

 

[Calin]

There are tasks that are very little dependent of interprocessor latency/bandwidth (like SETI@home or ray tracing). However, there are other (very interesting from usability standpoint) tasks that are mainly lots of interdependent processes working on a huge piece of shared memory. This is where latency and bandwidth limitations kick in.
Meteo is pretty much scalable (as weather in a zone affects only zones close to it) - a pass on the process changes everything, but the changes must be propagated only to nodes/zones close to affected one.
However, simulating nuclear explosions (or nuclear stockpile aging) is pretty much unscalable, as neutrons/positrons have a bigger probability to leave the area after they are generated than to hit something close. Every pass on the process changes everything, and all those changes must be propagated to every node/zone

 

[CTho9305]

Originally posted by: AluminumStudios

Originally posted by: CTho9305
A 600MHz Athlon can do 2.4gflops, so you'd need 417 of them to get a teraflop. Of course, since you can get 2.4GHz Athlons, 105 would suffice, but that's the absolute peak theoretical performance. You'll never see it in the real world. The easy way to get a teraflop (though not general purpose) would be to put together 5 boxes with top-of-the-line video cards (which can do >200gflops each).

I'm afraid parallel processing doesn't scale that well or work as simply as you've stated. If you go look at #401 on the Top500 computer list (http://www.top500.org) you'll see it's just shy of a terraflop with 900 2.0 GHz Opteron processors.

Memory bandwidth, inter-cpu communication, and other latencies really severly cut performance as you scale upswards in CPUs. Actual performance is MUCH LESS than the sum of what the CPUs are capable of.

See bolded sentence.

 

[BitByBit]

Originally posted by: AluminumStudios
If you read up on the challenges that Intel and AMD face in bumping up the clock speed the answer seems to be no - you can't have a 1 THz chip. The laws are physics are against you. The time it takes electrons to move from place to another plus the times it takes gates to open and close plus the capacinance, resistence, and current leakage issues would be pretty much insurmountable with electricity based chips.

I find it difficult to believe that some time in the future, such a speed won't be attained.
Regarding electron flow, the distance a signal can travel each clock cycle can be easily calculated.
At 1THz, we have:

F = 10^12Hz
T = 1 / F => T = 1 / 10^12 s. (This is the duration of each clock cycle.)

The distance the signal can travel in this time is given by:

d = vt

v = 3X10^8 m/s (Assuming the signal travels at light speed)

Thus, we get:

d = 3X10^8 / 10^12 = 3X10^-4 m
= .03cm

That is, the signal can travel 0.03cm each clock cycle.

According to my rudimentary knowledge of processor architecture, at this speed, it means no two pipeline stages can be more than 0.03cm apart, which to me, sounds like a rather large number when considering the infinitesimal distances one would expect to find within processor cores.

Perhaps the guy over at Chip-Architect could shed some light on this.

 

[f95toli]

Unfortunately, there are other problems as well. It is more or less impossible to clock silicon circuits above say 50 GHz; it has to do with the gap structure of silicon and is not related to the speed of the electrons.
GaAs and InP might work up to 100-200 GHz if we are lucky.

 

[Hyperlite]

there was just an article in the PC mag i got yesterday (so in effect its old news) that claimed a few scientists had managed to create a transistor that could operate at 603 Ghz.

 

[CTho9305]

IIRC, those transistors are BJTs, not MOSFETs. The downside of BJTs is that they consume power even when the output is switching, though they can be made to operate at extremely high speeds. Also, even a MOSFET that could switch 600 billion times per second won't get you a 600GHz chip - you need multiple gates worth of logic each clock cycle (each of which can require multiple transistors to switch), and have to worry about things like clock skew which further slow your cycle time.

One thing I've wondered about is the gain of these BJTs - if MOSFET gate leakage keeps growing, will there be a point where their static current becomes significant relative to BJTs?

 

[cquark]

Originally posted by: f95toli
I know of one such project. It was known as the "teraflop"-projekt and was based on superconducting RSFQ-technology. The research was done at Stony Brook (SUNY). Head of research was Konstantin Likharev.

Published work on RSFQ includes circuits running at speeds up to 770 GHz, so it seems they're approaching the TeraHertz goal. They produced their first RSFQ microprocessor in 2002, the Flux-1 which runs at 20 GHz.

Earlier posters mentioned teraflops, but even early RSFQ-based supercomputers will be in the petaflops range. A recent paper (2003) described plans to build a supercomputer consisting of 4096 RSFQ-based processors running in the 50-100 GHz range. http://www.iop.org/EJ/S/UNREG/EKtc3C13....5.2buV8iQ/abstract/0953-2048/16/12/010

 

[f95toli]

The 770 GHz circuits are very simple, i.e. flip-flops etc. The main problems are sill the fabrication technology (reproducibility) and how to make interconnects, so right now there is no way to scale this up to usefull circuits.
Circuits with maybe 10-20 0000 elements running at 60-70 GHz are possible and are actually being tested (I share lab-space with the RSFQ-group at my university).

The basic rule of superconducting technology is that one should not even try to do someting that can be done using semiconductors, this means that only circuits >100 GHz are interesting for applications; for slower speeds you can use e.g. InP and there is no way to compete.
Right now Japan is the only country with any long-time commitment to RSFQ-techology and their fabrication technology is already very impressive, we can probably expect DSP-type circuits running at 100-150 GHz within 5 years.

 

[linuxconvert]

I read in a computer magazine that soon they will be using diamonds with a microcoil of tungsten. and since diamonds have a higher heat tolerance, would it then be possible to get a 1THz proc?