Slashdot Mirror
IBM Creates World's Fastest Semiconductor Circuits
Todd Heidesch writes: "'IBM announced it has created the world's fastest semiconductor circuit, operating at speeds of over 110 GigaHertz (GHz) and processing an electrical signal in 4.3 trillionths of a second.'
IBM expects the new technology to be pumping out 100 gigabit/sec network switching chips by the end of the year (on an optimistic schedule, I presume)." dr_zeus contributes a link to this Reuters article running on Wired (also fairly thin) on the release, writing: "Granted, this isn't a PC chip, but one wonders how long it will be before we hear 'dude, you've got a 110GHz Dell!'"
At 110GHz, light travels less than 3mm in one clock cycle -- less than the width of the processor, I presume. And if it's accessing memory from a RAM chip 10cm away, it'll be waiting close to a hundred clock cycles to get anything back.
is in their ability to save power. From what IBM is saying, is that their chips can be run at say only 20 - 40 ghz and consume a hundred times less power then a chip built with todays processes. So you'll be able to get the same or more processing power out of these chips for less enegry.
At 110 ghz, a PHOTON only moves 2.7mm so figure that the actual signal propagation is like 2/3 the speed of that and you see that the signal can only travel 1.8mm in a clock. So, these chips are not going to be all that great for CPUs at 110 Ghz. Much better for signal processing likein routers or something.
That means ~1.29mm at C (speed of light), so about 0.9mm in reality. Wow, those better be some short circuit traces!
A SiGe process is an fudge to a bipolar technology process which is an addition to a more standard digital process. This means the devices are not your standard digital logic FET devices. The devices are most likely NPN vertical bipolar junction transistors, with the SiGe implant. The logic gates would then be standard complementary logic (CML) structures. Technology Description
Did you know that P4 has a couple of pipeline stages that do nothing but propagate signal? (yes, they pipelined the wire ...)
The Raven
The Raven
Call me stupid, but why can't they use the same material in PCs to increase the chip speed? Are there some limitations/incompatibilities other than the comparitively slow speeds of memory and I/O (I guess we can all see why I never got very far in that EE major...)
First of all, the IBM transistors are not MOSFETs, the tiny switches used in CPU's and other logic-based circuitry. They are instead heterojunction bipolar transistors (HBTs). HBTs are lightning fast and can be used as low-noise amplifiers for high frequency signals, which makes them great for wireless and Gigabit optical communication applications, but they are relatively large compared to MOSFETs and so are not really suitable for making CPU's. (Notice that the IBM press release never mentions CPU applications, but instead focuses on 100 Gigabit optical communications networks).
Now, you may wonder why SiGe can't be used to make super-fast MOSFETs. The main problem is that MOSFETs require a dielectric, such as SiO2 to act as an insulating layer between the "gate" and the channel. However, attempting to grow a layer of SiO2 on SiGe results in separation of the Ge from the Si, ultimately causing device failure. Currently, people are trying to find ways to deposit new dielectrics with higher dielectric constants, such as ZrO2, to replace the SiO2. Once this is acheived it may be possible to put such a material onto SiGe to allow creation of a MOSFET using this technology. However, development of such high-k dielectric technology is probably 3-4 years away and adaptation of this to SiGe will be a few more years beyond that, so don't expect SiGe-based CPU's anytime soon.
One last thing. I don't understand why IBM gets all the press. Motorola announced 110 GHz HBTs last October. IBM is really not as far ahead of the curve as they would like you to believe.
You're not limitted by how fast an electron can move, exactly. In fact, electrons move VERY slowly in common situations - the drift velocity in home wiring can be several feet per *second*.
When you shove a few extra electrons in one end of a wire, the charge pushes a few electrons that were already IN the wire down a little. And they push some down a little, and they push some down a little. Just like standing in a tight line at the movies, and shoving the guy in front of you - it takes a little bit of time to propagate all the way down.
So the real question is "If I shove an electron in this end of the conductor, how long before I get one out the other end?" The two things that determine that are (1) the nature of the conductor, and (2) the length of the conductor. By keeping the amount of circuitry on the IC very, very small (which they assuredly did), the propagation time from one end to the other drops proportionately.
However, even beyond just making the die smaller, they are working on making materials propagate the electrical charge more quickly - recently, someone (probably IBM) showed that by using a stressed crystalline lattice, they could significantly decrease the amount of time it took to propagate from one end to the other.
steve
Oh, you're not stuck, you're just unable to let go of the onion rings.
I'd still like to have even that modest potential, which would allow MAC (Multiply ACcumulate) operations at 10MSPS, for digital radio projects, etc. If you decided you need a different feature, just reprogram the fabric.
With today's technology, I don't see why you couldn't have a 4096x4096 grid with 4 way interconnects, running with at least a 1 GHz clock. This could do real time FFT, etc, straight from RF to anything. You could implement a crossbar switch in software for at least 32 streams (being conservative) at the clock rate, in software, with plenty of capacity to spare.
Processor fabric is a powerful concept, but Intel will never implement it, it's too much of a threat to them and their Von Neuman architecture. Someone else has to do it.
--Mike--
Okay, so I'm a real EE who design in IBM SiGe processes 5HP and 6HP.
1) IBM did demonstrate a ring oscillator.
2) These are IBMs latest SiGe HBT transistors, targetted for the "8HP" process. At present, 5HP and 6HP are in production and producing ICs - a lot of GSM cell phones will have IBM silicon in them. 7HP is coming on line.
3) Yup - these process are not directly for PC processors. The processes are targetted at RF, electro-optical, high speed data etc. They have SiGe transistors and CMOS. The SiGe is typcially used as a front-end, e.g. 10gigabit mutliplexers and laser driver/demultiplexors and diode detectors for optical links and the CMOS does the back end processing - e.g. line equalization etc.
In addition, this is not the fastest semiconductor circuit. For many years people have been using semiconductors at tera-Hz for microwave stuff (granted maybe not ring oscillators but certainly parametric-active amplifiers). I worked on 94GHz radar systems over 10yrs ago that used active semiconductors (IMPATT and Gunn GaAs oscillators).
That's why copper wires were important - they reduced R. C on the other hand is a different matter - for years and years (untill about 3-4 years ago) no-one cared about the capacitance of wires - because they were usually small compared with the capacitance of gates and the ratios tended to scale down as device features scaled down - everything got faster together
To make matters worse many of our CAD tools have untill quite recently made statistical guesses about wire capacitance which worked OK during things like synthesis (compiling to gates) when the wire capacitance was a small part of the equation, now it does matter and means the the whole structure of synthesis tools will have to change to perform combined synthesis and layout operations in order to create optimal circuits