Original Post
Memristors
This particular discussion is clearly oriented towards the technically minded, but I encourage you now if your at all interested in overpowering computers to the point of putting a full Supper computer cluster on a 1CM^2 chip, then it's worth reading!
So you may or may not have herd about memristor technology, and thus I will section this out for those that want to skip the introduction, heres the topics listing:
1: What is a Memristor?
A memristor is a fundamental component of electronics, one of the four fundamental components, (Resistor, Capacitor, Inductor, Memristor) it is eventually, just a resistor with memory. So what does that mean? Well resistance is just the measure of how much power (voltage) you need to drive electrons with, to push that at whatever speed (current). The interesting part is this devices resistance can change, and this change can be controlled via a voltage field. Still with me? Doesn't really matter, the good news is this device is fully capable of all logical operations, and also store digital information.
You want to know how this is possible? OK fine here we go:
(A short disclaimer, I've read on this topic, and watched video series, I'm also studying electrical engineering, and computer science, but i'm not an official on the subject, seek a real source.)
Titaniumoxide compounds are used, first one takes away some oxygen atoms and this leaves bubble like gaps in the remaining compounds, these gaps respond to fields of charge, and so you can move them to one side of your choice.
So why do you want to move the gaps? well by moving atoms around you can change the total resistance in the component, and with no field active it will maintain the same resistance, for as long as you leave it. Thus you get the memristor effect.
2: What can we use memristors for?
Memristors have innumerable uses, first a short list of important features:
So where dose this leave us, lets address storage first, as computer storage space is a relevant topic nowadays.
Storage:
HDD's, DRAM, SSD's of PCM and NAND Flash, are all suffering from major draw backs, I wont get into these but the memristor fits nicely into the markets center place, succeeding where everyone else fails. Sure the memristor isn't as fast as RAM, but on the other hand memristors are smaller, we'll get into why thats so important next:
The size of a memristor is impressively small, by my calculations I estimate you can fit 17,000 stacked against each other behind a single average human hair, width ways. This is important for a number of physics reasons, but we are mainly interested in the fact that it means the following:
A) By using such small memristors, we can create a 1CM^2 CMOS powered chip that multiplexes and demultiplexes information into a 4 dimensional array of memristors.
B) This is done mainly by having your 2D array of memristors, stacking them to get a 3D array, then using multiple states per memristor, to make it a 4D array. (cool eh?)
C) We can get about 1000 layers on a 1CM^2 chip, with that much CMOS.
D) That many memristors adds up to about 1TB per layer. Yes one whole Tera-byte, 2^43 bits (I believe (I'm sleepy k?)).
This results in about 1 Peta-byte of information on a 1cm by 1cm chip. Apple say that you can fit about 7000mins on an GB MP3 player, so what about a 1PB MP3 player? Well after doing the math you can have up to: 1750.42735 YEARS of music. And remember this is a 1CM^2 chip! RAM takes up a good 10x3 CM as a random guess (I've never measured a DRAM stick size).
More advanced storage:
But wait theres more!
People have now been thinking about using this with CPU's and putting memristors onto the cache; they wont take up much power, will provide enough cache to forever ignore RAM, wont cost much, the data rate on a CPU's cache far exceeds that of ram due to the bandwidth, and memristors can provide such a bandwidth as well.
This would mean you could have a non volatile, hard drive, cache, RAM, clusterfuck on a CPU. It would be faster than current RAM due to the bandwidth, it will be so much faster than hard drives it's funny, and will be large enough to be viable as primary storage. Ever wanted to load windows in a few micro seconds?
A few more brief examples:
There are a whole bunch more topics here that deserve coverage, but I simply don't have the time to write about them, so I'll mention them:
Logic: Memristors are incapable of doing a NOT operation, however while some people might think this is a rather large blow, as the NOT is required for NAND logic, and thus all electrical logic, your missing out on a different fundamental logic paradigm; material implication.
What is material implication? No idea, looking into it! However mathematicians studying logic note that material implication is many times more efficient than NAND based logic, and a scientist at HP Labs, discovered that while programing material implication logic compilers in C, and comparing them to NAND logic compilers, the compiled code for the same logical deductions turned out to be three times smaller in material implication than NAND.
So why is this good, well the Memristor device is 3nm vs the transistor at 32nm currently for CPU's, and the Memristor has more than two states, and the memristor is capable of more efficient logic, with an estimated three times less components needed for the same results. New stupidly power efficient CPU's anyone? (That was less brief than I expected)
Neural Networks:
And now I step way out of my depth, infact I drowned a few paragraphs ago, but really I can sum this up quickly:
The human brain's neurons are similar to transistors, the synapses are memristors, yes, are memristors.
IBM recently simulated a cat brain using a huge supper computer.
HP designed a chip that can do the same thing that is just over 1cm x 1cm big.
Pwned.
3: Problems that are faced:
Memristors aren't as quick as transistors.
Memristors still only have a life time that only just exceeds flash, it's not yet the clear winner in life time. (however it is believed this will be over come)
Memristors are basically owned by HP, so we have to wait for HP to do the work.
Memristors are vastly different to transistors, this presents new design problems, likely to get a computer at home to work on memristor tech (partly or fully doesn't matter, we just use whatever suits the situation best) the motherboard, RAM, CPU, GPU, Bus', and so on and so forth will need a complete redesign, with new protocols. We can start on just the protocols and buses, introducing the memristor hard drives, and NVDRAM (non-volatile DRAM). But this will take time to adapt.
Others I can't think of right now.
4: What we have today:
HP have started manufacturing Memristor chips, and say NVDRAM is scheduled for 2013. Personally I think thats a touch ambitious, I'd say 14/15, and mostly I want so see some SATA6Gb/s controllers on these babies, so I can start using good hard drivers.
So torifag's, whats your take, and how impressed with HP's scientists are you?
Also if you liked this post but can't think of anything to say just derep me for saying everything there was then posting it in discussion, seems a tad hypocritical? Oh well.
Main source: http://www.youtube.com/watch?v=bKGhvKyjgLY
So you may or may not have herd about memristor technology, and thus I will section this out for those that want to skip the introduction, heres the topics listing:
- What is a memristor? (in laymen terms)
- What can we use memristors for?
- Problems that are faced.
- What we have today.
1: What is a Memristor?
A memristor is a fundamental component of electronics, one of the four fundamental components, (Resistor, Capacitor, Inductor, Memristor) it is eventually, just a resistor with memory. So what does that mean? Well resistance is just the measure of how much power (voltage) you need to drive electrons with, to push that at whatever speed (current). The interesting part is this devices resistance can change, and this change can be controlled via a voltage field. Still with me? Doesn't really matter, the good news is this device is fully capable of all logical operations, and also store digital information.
You want to know how this is possible? OK fine here we go:
(A short disclaimer, I've read on this topic, and watched video series, I'm also studying electrical engineering, and computer science, but i'm not an official on the subject, seek a real source.)
Titaniumoxide compounds are used, first one takes away some oxygen atoms and this leaves bubble like gaps in the remaining compounds, these gaps respond to fields of charge, and so you can move them to one side of your choice.
So why do you want to move the gaps? well by moving atoms around you can change the total resistance in the component, and with no field active it will maintain the same resistance, for as long as you leave it. Thus you get the memristor effect.
2: What can we use memristors for?
Memristors have innumerable uses, first a short list of important features:
- Memristors works best when small, and are currently down to the 3nm scale.
- Memristors have a switching speed of about 1ns max(ish) without degrading total life time expectancy intolerably.
- Memristors can be switched at around 1/10th the speed of the RAM in your computer (DRAM)
- Memristors are 3nm's, transistors are around 32nm's, and flash is dropping to about 20nm. Smaller is better here.
- Memristors unlike DRAM are non-volitile, and won't loose thier resistance.
- Memristors have a new architecture, which allows them to be stacked one atop the other.
- Memristors can store more than two states.
- Memristors are cheep as s***.
- Memristors have very very small power consumption.
So where dose this leave us, lets address storage first, as computer storage space is a relevant topic nowadays.
Storage:
HDD's, DRAM, SSD's of PCM and NAND Flash, are all suffering from major draw backs, I wont get into these but the memristor fits nicely into the markets center place, succeeding where everyone else fails. Sure the memristor isn't as fast as RAM, but on the other hand memristors are smaller, we'll get into why thats so important next:
The size of a memristor is impressively small, by my calculations I estimate you can fit 17,000 stacked against each other behind a single average human hair, width ways. This is important for a number of physics reasons, but we are mainly interested in the fact that it means the following:
A) By using such small memristors, we can create a 1CM^2 CMOS powered chip that multiplexes and demultiplexes information into a 4 dimensional array of memristors.
B) This is done mainly by having your 2D array of memristors, stacking them to get a 3D array, then using multiple states per memristor, to make it a 4D array. (cool eh?)
C) We can get about 1000 layers on a 1CM^2 chip, with that much CMOS.
D) That many memristors adds up to about 1TB per layer. Yes one whole Tera-byte, 2^43 bits (I believe (I'm sleepy k?)).
This results in about 1 Peta-byte of information on a 1cm by 1cm chip. Apple say that you can fit about 7000mins on an GB MP3 player, so what about a 1PB MP3 player? Well after doing the math you can have up to: 1750.42735 YEARS of music. And remember this is a 1CM^2 chip! RAM takes up a good 10x3 CM as a random guess (I've never measured a DRAM stick size).
More advanced storage:
But wait theres more!
People have now been thinking about using this with CPU's and putting memristors onto the cache; they wont take up much power, will provide enough cache to forever ignore RAM, wont cost much, the data rate on a CPU's cache far exceeds that of ram due to the bandwidth, and memristors can provide such a bandwidth as well.
This would mean you could have a non volatile, hard drive, cache, RAM, clusterfuck on a CPU. It would be faster than current RAM due to the bandwidth, it will be so much faster than hard drives it's funny, and will be large enough to be viable as primary storage. Ever wanted to load windows in a few micro seconds?
A few more brief examples:
There are a whole bunch more topics here that deserve coverage, but I simply don't have the time to write about them, so I'll mention them:
Logic: Memristors are incapable of doing a NOT operation, however while some people might think this is a rather large blow, as the NOT is required for NAND logic, and thus all electrical logic, your missing out on a different fundamental logic paradigm; material implication.
What is material implication? No idea, looking into it! However mathematicians studying logic note that material implication is many times more efficient than NAND based logic, and a scientist at HP Labs, discovered that while programing material implication logic compilers in C, and comparing them to NAND logic compilers, the compiled code for the same logical deductions turned out to be three times smaller in material implication than NAND.
So why is this good, well the Memristor device is 3nm vs the transistor at 32nm currently for CPU's, and the Memristor has more than two states, and the memristor is capable of more efficient logic, with an estimated three times less components needed for the same results. New stupidly power efficient CPU's anyone? (That was less brief than I expected)
Neural Networks:
And now I step way out of my depth, infact I drowned a few paragraphs ago, but really I can sum this up quickly:
The human brain's neurons are similar to transistors, the synapses are memristors, yes, are memristors.
IBM recently simulated a cat brain using a huge supper computer.
HP designed a chip that can do the same thing that is just over 1cm x 1cm big.
Pwned.
3: Problems that are faced:
Memristors aren't as quick as transistors.
Memristors still only have a life time that only just exceeds flash, it's not yet the clear winner in life time. (however it is believed this will be over come)
Memristors are basically owned by HP, so we have to wait for HP to do the work.
Memristors are vastly different to transistors, this presents new design problems, likely to get a computer at home to work on memristor tech (partly or fully doesn't matter, we just use whatever suits the situation best) the motherboard, RAM, CPU, GPU, Bus', and so on and so forth will need a complete redesign, with new protocols. We can start on just the protocols and buses, introducing the memristor hard drives, and NVDRAM (non-volatile DRAM). But this will take time to adapt.
Others I can't think of right now.
4: What we have today:
HP have started manufacturing Memristor chips, and say NVDRAM is scheduled for 2013. Personally I think thats a touch ambitious, I'd say 14/15, and mostly I want so see some SATA6Gb/s controllers on these babies, so I can start using good hard drivers.
So torifag's, whats your take, and how impressed with HP's scientists are you?
Also if you liked this post but can't think of anything to say just derep me for saying everything there was then posting it in discussion, seems a tad hypocritical? Oh well.
Main source: http://www.youtube.com/watch?v=bKGhvKyjgLY
Last edited by Vox; Jun 11, 2010 at 12:17 PM.
I think this will be awesome, and I'm very impressed. As far as I know, standard memory is two NAND(or NOR, I forget) gates per bit. The simplification here will be wonderful.
i have a totally post modern tattoo of a scalene triangle.
<DeadorK> fair maiden<DeadorK> if the cum is going to be in your mouth
<DeadorK> it shall be in mine as well
Reading the first paragraph right now. So really it's just a resistor that can be altered by changing its chemical makeup? Doesn't sound commercially viable or particularly revolutionary. Will edit when finished reading, there's probably something I'm missing. EDIT: Oshite, didn't realize that they were so small or easily manipulated. Well this is totally awesome, too bad HP's going to be the only one developing it. Oh, and a question. How are they able to control the compound's oxygen distribution at such a tiny level so uniformly?
Last edited by Boredpayne; Jun 11, 2010 at 11:19 AM.
Long story short the degree at which you can control memristors increases the smaller it gets.
But more importantly the guy did the maths, and can specifically predict states based on some very good calculations.
[link]
Added that link to the bottom of main post.
[EDIT]
I really want to see a CPU cache used as a hard-drive personally, some CPU's have a data rate to their cache of around 100GB/s (albeit that's the intel i7 980X) and memristors, while slower could easily have a much bigger bandwidth, thus keeping up with this speed, or even surpassing it.
Plus you'll only need a small amount of the chips space for this, no more than is currently used, and possibly less. And you could have a 1TB cache! (That you could read in 10s!)
How big of a step is that? Well the intel i7 980X has a 12MB cache, and in a smaller space by stacking the memristors we could easily fit a few terabytes.
Victory.
But more importantly the guy did the maths, and can specifically predict states based on some very good calculations.
[link]
Added that link to the bottom of main post.
[EDIT]
I really want to see a CPU cache used as a hard-drive personally, some CPU's have a data rate to their cache of around 100GB/s (albeit that's the intel i7 980X) and memristors, while slower could easily have a much bigger bandwidth, thus keeping up with this speed, or even surpassing it.
Plus you'll only need a small amount of the chips space for this, no more than is currently used, and possibly less. And you could have a 1TB cache! (That you could read in 10s!)
How big of a step is that? Well the intel i7 980X has a 12MB cache, and in a smaller space by stacking the memristors we could easily fit a few terabytes.
Victory.
Last edited by Vox; Jun 11, 2010 at 02:23 PM.
