160GHz into the future

[utopia]

Cool Project

----------

This is an extremely interesting article, I strongly suggest you all check it out. This is one of those "Elite" posts

---------



<< A multiprocessor COOL system based on superconductor Rapid Single-Flux-Quantum (RSFQ) technology is being developed at Stony Brook within the framework >>



Looks like they are going beyond CMOS (Complementary Metal Oxide Semiconductor) technology, and implementing a whole new (and unknown) complex process.



<< to achieve petaflops-scale performance (~1015 floating-point operations per second) by 2007 >>



We are not talking about a system that can yield up to a PetaFlop level! Amazing, but still to be achieved...



<< . The concept system has multiple levels of distributed memory, namely holographic HRAM, semiconductor SRAM and DRAM, and cryo-memory (CRAM) >>



HRAM has been rumored to be the next generation-advanced technology, which will be able to operate at frequencies of GHz's, and have access times of as low as 0.5ns'. IBM and MIT are working on this technology as well.



<< RSFQ Superconductor Processors (SPELLs) operating at the temperature of liquid helium (4.2 Kelvin). >>



This is the high point of the article... I seriously doubt that 4K (-269C) is possible at this time. Liquid helium is able to bring the temperature to such extremes, but for a very short time.



<< The peak CRAM bandwidth is eight 64-bit words per 30-ps cycle, i.e., ~ 2 TB/s >>



2 Terabytes (1TB=1024GB) of memory bandwidth throughoutput! Hardly believable, but is possible to be achieved with such technologies as multi-stage/channel DRAM. Think multi-channel DDR...



<< MSU can work at a much higher rate of up to 160 GHz >>



-Moore's law might be proven wrong, but still under a huge question.



<< We have found that the COOL system with 4,096 SPELL processors and 4 GB CRAM can achieve near-petaflops performance, while occupying a physical space of ~ 0. 6 m3 and dissipating ~400-500 Watts at 4 Kelvin (~150 KW at room temperature). >>



I doubt that this machine will be able to fit into this space (less than 1 m3!), 500 watts sound also unreal.

I would like to know how they are planning to cycle Liquid Helium, which hasn?t really been achieved while keeping it at under 50 Kelvin in a closed-loop system.


Good read, but vaporware for now...

 

[utopia]

Bump

 

[Train]

drool...

 

[utopia]

Hell, plug in 20 GF2 Ultra's into this baby, and Q3 never looked better!

 

[downhiller80]

I think I'm right in thinking that if the CPU('s) are superconducting at 4K then so long as they are kept at 4K they will generate little, if any heat themselves.

Of course the ambient room temp will still cause this stuff to evaporate, but what if they were to contain it and pressurise it?

This would keep it at a low temp wouldn't it?

Seb

 

[utopia]

<< if the CPU('s) are superconducting at 4K then so long as they are kept at 4K they will generate little, if any heat themselves. >>



At this state the core becomes very efficient (in theory), and the resistance is minimal. The core will still generate all of the original heat (ex. 120Watts), but that heat will be transferred to the coldplate of the cooling device (or direct gas). Remember that heat is transferred from hot to cold.




<< Of course the ambient room temp will still cause this stuff to evaporate, but what if they were to contain it and pressurise it? >>



You are thinking Freon systems now, but that?s what I was talking about when I mentioned the closed-loop system. The Helium will be inside a compressor, and will be cycled through the system by a special pump. The problem is that helium as well as nitrogen heats up fairly quickly, and it's virtually impossible to boil it back (many times) into 50K/4K state. This is not a problem with Freon gas, (Think Kryotech, Asetek) because it is easily cooled with just a simple radiator and passive cooling (fan).

 

[utopia]

Lets get some people in here.. Where is PM?

 

[Adul]

wonder how rc5 would crack on that.

 

[Pretender]

That'd get rc5 done in a few weeks at most prolly.

 

[andri]

On quantum computers, encryption and decryption run at the same speed
So, the right rc5 answer would be available in ~0.1 seconds.

Think about it No more rc5-like projects, as our current methods of encryption arent secure on quantum computers - decrypting it is just as easy.

Heck, almost all computations will be done in *one* clock tick...

 

[utopia]

Then we would need RC5's with like 50000000000000000000000000000000000000000000000000000000000000000 blocks!

 

[noxipoo]

don't think you need 20 GF2 ultras Utopia, the CPUs can do much better i would think

 

[utopia]

Well, i was joking. Better in what?

 

[Haervii]

Peta Flops? PETA FLOPS!!!!! PEETTTAAA FLLLOPSSS!!!!!! HOLY SMOKES, NOT PETA FLOPS!!!!!!!!! (GASP)

 

[pm]

Utopia: I'm not sure what I can add to this discussion. I'm definitely no expert in cryogenics, and I really don't know that much more than average Scientific American reader about superconducting circuitry. I'm vaguely familiar with RSFQ's - they are superconducting circuits that use multiple superconducting Josephson junctions to control data through inductance. The ironic thing about them is that &quot;normal&quot; silicon designers primarily pay attention to capacitive effects and inductance is a non-issue due to the low magnetic susceptibility (right term?) of silicon. RSFQ's work primarily due to inductive effects with capacitance being a secondary issue. It's an interesting switch.

I'm not sure how realistic this device is for future product, but it is exciting from a research perspective. I wasn't aware that anyone had a working prototype of an HRAM memory device, and the CRAM bandwidth is extremely impressive.

Edit: the last line in the first paragraph was an unintentional pun in case you are wondering.

 

[utopia]

Well, i just wanted to find out if you think that Terabytes/sec, HRAM, and Petaflops are realistic... You do work with 64bit Intel CPU's after all!

 

[Mr.IncrediblyBored]

<< Edit: the last line in the first paragraph was an unintentional pun in case you are wondering. >>

*wooooosh* The sound of the pun going right over KB's head.

 

[pm]

Well, who am I to contest their numbers? If they say that's the bandwidth and those are the performance specs, then I'm not really in a position to say &quot;you are wrong.&quot; They say that they can get 512 bits transferred per 30ps clock period (which is ~2TB/s). I don't have any evidence that they are making it up. The only thing that struck me about the memory bandwidth is that they don't explain how fast the &quot;PNET&quot; bus between the memory and the CPUs is capable of running. They certainly seem to be on the cutting edge of computing: multithreading, multistreaming, hiding latency through prefetching, etc. It sounds impressive, but they don't give enough details to determine any holes or flaws - especially noteworthy is that they don't really discuss the computing architecture of the SPELL processors. It really hard to judge much of anything from a paper that sounds like an abstract for a real paper, except to note that they are pushing the cutting edge and it sounds like exciting research to me.

Edit: the pun was on the words &quot;an interesting switch&quot; - which was intended to mean that it's a change for designers to have to consider inductance, but, since transistors are often called switches, it could also refer to the fact that an RSFQ transistor is &quot;an interesting switch&quot;. It's a bad pun, and it was accidental. I noticed it when I read through my post after I'd hit 'post'.

 

[Mr.IncrediblyBored]

Thanks for clearing that up, PM.

(I get it now )

 

[utopia]

I am a bit skeptical on the whole &quot;4K&quot; part. It?s just unreal at this moment. I believe something like 150K or 75K, but not 4K.

 

[Gatsby]

Now the big question is..

How fast can it crack rc5?

Gatsby

 

[office boy]

This fast..
[Oct 05 23:55:40 UTC] 1 cruncher has been started.
. Done
[Oct 05 23:55:42 UTC] Conratulations you have found the key!

Maybe faster...
Read andri's post.