Slashdot Mirror
New Alternatives To Silicon May Increase Chip Speeds By Orders of Magnitude.
First time accepted submitter Consistent1 writes "A paywalled article in the "Nature Materials" journal describes the use of Magnetite to achieve ultra fast electronic switching, albeit, at the moment, only at extremely low temperatures. According to a story on Quartz, the team, led by Dr. Hermann Dürr from the Stanford Institute for Materials and Energy Sciences hopes 'to continue the experiment with materials that can operate at room temperature. One possibility is vanadium dioxide.' Chips utilizing this technology may operate at clock cycles thousands of times faster than the silicon-based chips used today."
I taught we already had gallium-arsenide transistors. The problem is cost as it is reserved for application where power enveloppe is very thin (earing aids) and switching speed is critical (telecom equipment).
Tomorrow is another day...
If this technology became mainstream, I'd bet my IBM Model M13 that people would still try to overclock the shit out of it.
You do understand that somebody has to do groundwork before anything can be made in large scale. Even first silicon transistors where originally just proof of concepts until engineers where able to make manufacturing process around it.
I thought one of the main issues with increasing clockspeeds on processors besides heat is also the latency. at 3 Ghz a signal can only travel 10 cm anymore, and processors already have stages in their pipelines just to get the signals around. So going 1000 fasters would have to mean some major changes in how processors work i guess? since having your signal only travel 0.1 mm per clock pulse makes it rather hard to get the data around...
Fucking slashdot, with its lack of support for basic unicode. What is this? 1996?
Back in the days, when slashdot...
That's a bit of an obvious troll coming from someone with a seven digit UID... :p
"Convictions are more dangerous enemies of truth than lies."
The first working Silicon transistor was 1954 and worked at room temperature. The first microprocessors were in the late '70s. It's great that people are working on other materials for transistors, but it's a very long road from 'works in the lab' to 'ships in a mobile phone'. 20 years is not unusual.
I am TheRaven on Soylent News
No, the clock signal needs to time between two connecting flip flops nothing more. It's extremely common (i.e. it's about 5% of my job) to have to change the design in order to achieve this local clocking requirement.
That's without having multiple asynchronous clocks on a single chip.
Or asynchronous logic
Even when you need to do very long paths it's called a clock tree for a reason you can have a 1GHz clock that takes several ns to get from its source PLL to its destination flop because the delay through the tree to all the leaf nodes is matched. that is a 1ns period clock can take 4ns to get from the source to the destination, and that's all fine because as long as it's the same 4ns... ;-)
Now things get harder when different bits of the chip have silicon that runs at different speeds so you can't balance the tree like you'd like to, but that's what makes this job interesting
"The weirdest thing about a mind, is that every answer that you find, is the basis of a brand new cliche" -
20-30 years seems to be a good rule of thumb. So if you want to know what the promising technologies of the next decade will be you should look at what has been done in the lab in the late '80s early '90s. (FDM 3D printing seems to be right on the mark, and if the Oculus Rift thing pans out VR will be too. Looking at stuff from the late '90s, electric cars will have to wait another decade to get mass adoption. LED lighting is ahead of schedule. Decent adoption rates a mere 20 years after the first superbright blue LED was demonstrated by Shuji Nakamura).
Also, a clock signal is a single-bit signal. You can use a wide interconnect for distributing it over large distances in the higher levels of the tree, making it much faster compared to the local interconnects. That makes it somewhat less of an issue than is the case with long-range data interconnects, which are parallel (or did they switch to serial lines even on-chip?), therefore have to use narrower interconnects, therefore are slower.
Ezekiel 23:20
Why are you posting here? Why bother?
Do yourself and everyone else a favor. Go away and leave the rest of us alone. We're better off without you, The only person who enjoys your whining is you. Stop it now.
Why is Snark Required?
How much energy it takes to switch 0/1 states? What voltage? As I am not in the field, it would take me too much time to extract this information from the article (what is "trimeron annihilation" and how/does it relate the classical hole-electron recombination?).
I assume that it is possible to be 1000 faster only if it takes considerably less energy to switch states. It means that even if the latency constrains the speed, it would still produce less heat and will allow simpler clock/power lines.
As I understand it, one of the major factors that slow the speed of today's electronics is power. Be it in the form of routing constraints (possibly wider metal lines and possibly wider minimum distance between them), power dissipation, battery capacity in mobile devices, or cooling in servers, all are constrained by power. If this technology can lower power requirements then there will be a significant speed-up either in the form of more cores on a chip, or newer computation models that work better with deeper pipelining or with wider SIMD operations.
Another potential advantage of the fast switching is that it enables or enhances other computing models. Maybe we will move farther away from a pure CPU programming model to an FPGA/CPU hybrid programming. It's time to brush up your VHDL/Verilog capabilities, or to teach your pet language (compiler/interpreter/JIT) how to emit an efficient HDL. The advantage of FPGA programming is that you can define your own pipelines according to the computing task at hand. Another thing to consider is that with these switching-speeds it could be profitable to time-share an FPGA. Finally, it may be possible to reprogram an FPGA in less than a second.
IIRC, making P-type material was easy doping with boron, and someone had finally come up with a way to make n-type material.
In addition, around that time there were two or three startups looking to manufacture diamonds using various -cheaper- processes. The combination of these things was supposes to give is diamond based chips that, due to the incredible heat resistance of diamond, could tolerate much more heat and hence higher clock cycles.
Does anyone know where this went?
Silence is a state of mime.
Does this mean I should stop having my dwarves smelt it into iron bars?
--- Math illiteracy affects 8 out of every 5 people.
Well, let's see. The Solar System weighs on the order of 10^30 kg. That's 2^100 kg. There's 2^86 atoms in a kilogram of hydrogen. That's only 2^186 hydrogens in our solar system, if its whole mass was hydrogen. You seem to be right - iterating through 2^256 is quite unfeasible.
Assuming iteration speed of 2^32/second, given 2^24 seconds per year, and a billion PCs worldwide (2^30), we could "crunch" only a space of 2^86. Our current resources are about a factor of 2^170 too small :)
A successful API design takes a mixture of software design and pedagogy.
If you want to have 1000 times shorter cycles, you need a 1000 times smaller chip.
Lets examine this..
The 80386 used a 1500 nanometer process. We are now playing with 22 nanometer parts (transistors that are 68 times smaller in length.)
The most common speed of the 80386 was 33 MHz, and the most common speed of a modern computer (according to the admittedly biased Valve Hardware Survey) is ~2500 MHz.
~2500 / 33 = ~75
So in practice what you are saying is clearly within an acceptable margin of true, but is perhaps not clearly stated (you need a 1000 times smaller process, not a 1000 times smaller chip!)
This does also show that the diminishing returns of higher clock speeds are likely real. If you want higher clock speeds without a smaller process size then you need a longer pipeline and thus higher instruction latencies, defeating a large chunk of the benefit of the higher clock speed.
However, for special purpose architectures (perhaps GPU's) with different use cases (where a deep pipeline doesnt have as many downsides), then higher clock speeds could be a big benefit even without a smaller process size.
"His name was James Damore."
Of course, most of the delay that limits clock speeds now is in the interconnect and not the switching devices. We're already using copper conductors and low-K dielectrics, so the next step is going to have to be superconducting interconnects.
Until then, it's mostly a laboratory curiousity.
Lacking <sarcasm> tags,
That's not because vanadium is rare but because silicon is absurdly abundant; there's more vanadium than chlorine, lithium, cobalt, copper...
I really doubt scarcity is an issue here.
No kidding!!! What do you say at this point?
There's 2^86 atoms in a kilogram of hydrogen.
Hmm, there are 6x10^{23} atoms of hydrogen in a gram of hydrogen, so that would make it 6x10^{26} hydrogen atoms in a kilogram of hydrogen.
http://en.wikipedia.org/wiki/Avogadro's_number