Big Blue demos 100GHz chip IBM reseachers have made a breakthrough in the development of ultra-high-speed transistor design, creating a 100GHz graphene-based wafer-scale device. And that's just for starters. The transistor that the researchers have developed is a relatively large one, with a gate length of 240 nanometers - speeds should increase as the …
If I understand this correctly, the -3db frequency is meant here. This means that a sinus input of 100Ghz achieves an amplification of -3db.
As a useful digital signal of n Hz needs harmonious frequencies of about 7n to 10n Hz, this means that we can expect digital circuits of 10 to 14 GHz operating frequency. That certainly is an improvement over those 5 GHz we already have, but not really dramatic as compared to current CMOS technology.
Actually for amplifiers you typicall give the Gain Bandwidth product, which is about the frequency you reach amplification of 1. For "boxed" amplifiers where you do have the 3dB frequency. It's 3dB lower than the normal signal amplification. This would still be OK for digital circuits.
Further than that, they are talking about the switching speed for a single transistor, presumably driving no load (they don't say if this is nmos or pmos either). A very very heavily pipelined processor might have logic about 8 fanout of four inverters deep, plus pulse latched registers would bring you to about 10 gates deep, so if the 100 GHz was for the digital range of the transistor, the system clock would be maybe 10 GHz, probably less once the effect of load is included. I'm not good with the analog side of things, so I'll bow to joeuro above and conclude that should be cut in 10 again, so maybe 1GHz system clock with the current generation. This jibes with the statement that these are 4x as fast as a conventional transistors with a similar gate length (40nm refers to the gate length after some depletion, so this is about 5-6 times as big as the bleeding edge, and by rough calculation, about 1/2 the speed).
All that said, this seems like some very cool technology-- IBM's hardware is always very impressive. I hope they succeed in bringing it down in size.
Beers to joeuro for quick thinking, and beers to the brainy bofffins.
to run Crysis? Erm, yes it will be - my pensionable 2.6GHz (single core!) Athlon64 manages to run it quite nicely when coupled to a (less elderly) 8800GTX gfx card. Now if your benchmark was CoD:WaW then that's a different story. <sigh>
Like most I'm wondering if a 10GHz processor will be fast enough to run the version of MS-Office that's out at the same time... (cheap shot I know, but it is a Monday morning).
Seriously though, IBM tech still manages to impress. Good on them.
Does this mean the software development industry can go back to the heady days of yore, producing bloatware as usual, secure in the knowledge that Moore's law will cover most, if not all, of their sins?
After all, cleaning up sloppy coding, ridding the application architecture of unnecessary middleware, or excessive levels of abstraction caused by the choice of language and development frameworks is real work. Boring, nasty business.
Without hardware resource constraints, we can allow developers to explore their "creative horizons" in interesting and sometimes exotic ways, without worry that the thing will hit the wall once you put a real-world peak user load on it.
I can almost hear the cheers already.
So when these new processors actually get made, all these super duper encryption technologies won't be worth diddly squat as they will get broken in hours if not minutes. Think proper security will be needed rather than hoping that it'll take a snoop years to find out the information.
No, they won't. You simply don't understand the issues and orders of magnitude involved here. Brute forcing, or even reduced key space brute forcing exploiting various loopholes would still require a geometric increase in throughput.
Obviously, with an NSA-sized budget, you could use an underground city-sized array, but all that money spent on researching how to bring some sort of reliability to the output of of multi-quibit quantum processors is the danger to crypto.
Increased CPU speed is only a linear improvement in breaking cyphers. So a 20x faster CPU will only take about 1/20th of the time as a current CPU. Since the better encryption technologies can resist a brute force attack for thousands of years, we're still talking decades or centuries
However, increased CPU speed can result in an exponential increase in cipher security, by allowing an increased key length. So assuming these CPUs aren't sold solely to hackers, encryption is hardly at risk.
If you want to break encryption, find the magic algorithm than can easily reverse what we believe to be one-way mathematical operations. Then don't patent it or share it with anyone.
As processing power increases, more becomes available for encryption as well as cryptanalysis. Longer, stronger keys become practical, so snoopers are no closer to easily breaking crypto. The only caveat is that any older data that is still sensitive needs to be re-encrypted with the newer schemes to remain safe.
What do you call proper security exactly? I thought it was good strong encryption, where people are encryption, although I fear that you think there is somthing other than encryption to most security. Breaches happen when the encryption is either not strong enough or not properly done (i.e. work arounds or errors)
The main problem most people don't get is that those 100 GHz is probably meant for small signals. Essentially you amplify some DC signal and add a bit of high frequency signal. This is how it's done in analogue electronics. And in fact there 12 GHz is already _cheap_ as you can see at 3 Euro LNCs for direct broadcast satellite reception.
In digital applications you usually want to switch a transistor all the way. If you don't do that in CMOS you will essentially short circuit your power supply and your chip will blow up. This did happen in some early CMOS designs.
There is actually one digital technology which aims to use the speeding benefits of analogue circuits is ECL Emitter-coupled logic. Those circuits can easily reach the multi gigahertz range, however a simple gate can take several milliwatts of power. They are usually used as dividers for frequency counters.
"So when these new processors actually get made, all these super duper encryption technologies won't be worth diddly squat as they will get broken in hours if not minutes."
That is plainly WRONG. If you would just brute-force a 128 bit symmetric cipher, you would have to test 2^128 keys. Even if you had this 15 GHz clock, you could at max test 2^33 keys per second. So you would still need 2^95 seconds to do that. Now that would be 1256154276291608599593 years....
Yeah, you are Uncle Sam and have 1 billion processors to work in parallel, but that still would reduce it to only 1256154276291 years. That is 1.2 Trillion.....
...would be good. Might lower too-high-in-recent-years prices on the latter's products too.
Although, to digress, I do resent it when we hear Ministers saying we all need to make better rather than cheaper - inspired as they clearly are by examples such as this, and oblivious of the effects of superinflated capital goods prices on anyone wishing to start up in business. We're not all fucking geniuses.
Is this a fair first approximation?
240 nM feature size versus 32 nM feature size => 56.25 = (240/32)^2. So 56.25 times fewer graphene transistors per unit area than silicon.
100GHz/3.33GHz => 30.03 time faster operation of a graphene transistor vs silicon.
At first blush this looses against silicon (30/56.25 < 1.0).
But if an IBM graphene CPU was made with 1/30 the number of transistors but simpler design needing less deep pipelines, less branch prediction, the graphene CPU could still be faster at, say, floating point or integer/logic but maybe not both, despite having many times fewer number of transistors.
I'm just a sysadmin and programmer, no kind of hardware guru. So I know I could be quite out to lunch...
First off, you're comparing individual transistor switch speed vs switching speed of a full pipe stage (probably 10-20 transistors deep, plus some pain from capacitive load and wire delay). The key bit is in the article where they say it is 250% the speed (2.5x) of a similarly sized silicon gate, so presumably the prototype is slower than a 40 or 32nm silicon gate. 100GHz number is attention grabbing, but not relevant to architectural speed.
Now, your idea about simpler architecture is not a bad one if the actual issue were a big fast gate vs a small slower gate, but there's two major problems. First, if you could make a significantly smaller core through the right choice of simplifying architectural decisions, with silicon gates, you could have more of those cores, or they could have wider execution paths (this was the driving principle behind the Cell's architecture). The other problem is that one of the major limitations on core speed is propagation delay along wires, particularly global interconnect. So, even if you have a very fast transistor, it still takes the same time for a signal to get from one side of the die to the other. This will be a killer if you can only fit a single core on your die. The more likely use case for big fast gates (assuming the processes were compatible) would be to mix them in sparingly along critical paths, and indeed the circuit guys already play all sorts of games with gate sizing (mainly for driving big capacitive loads quickly).
Yeah, I suspect it would run Doom :-) But it would cost about 100 times the price of a bottom end current x86 CPU! Almost certainly it would become a top dollar IBM blade server product CPU, just like the Cell. Grrr.
I think we got Doom on cell phones. Except maybe not the iPhone. You know, not wholesome enough.
Are ye missing something? They'll patent this, and lease the rights to use it, to everyone else, including Intel - same as they have been doing with all their inventions for 3 decades. Around 60-80% of the profits on a current Intel chip go straight Armonk, anyway, with none of the production overhead to worry about.
It's the same with high density disc drives, laser optics, RAM - a whole raft of the hard stuff that goes into anything, these days - from a mainframe to a E-book. There's not been a lot of computer, you could build, since the late 80s, that would even function without making use of some patented IBM technologies. Sooner or later, the money gets back to Armonk. Thats how you get to 'lose' control of a PC industry, and yet still somehow manage to remain the biggest IT company in the world.
We see a tentacle here, we see a tentacle over there - but somehow, no one ever seems to wonder if all the tentacles might actually joint together into something really big, that's right underneath them.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
