A few unanswered questions
Rob K.
I have a few general questions about Nanotechnology that I couldn't find
answered on the web.
1) Do we have a design for an assembler? I realize that building one is
still many years in the future, but is it too early to sketch out the
preliminary design?
2) Do we have a design yet for a Nanocomputer? What kind of instruction set
would it have? How much memory? What programming languages will assemblers
use?
3) What about networking? Trillions of assemblers are going to be working
together to accomplish a goal. Will TCP/IP work?
4) Finally, do we have computers powerful enough to simulate all of the
above? Has anyone tried it yet?
Will Ware
"Rob K." wrote:
> 1) Do we have a design for an assembler? I realize that building one is
> still many years in the future, but is it too early to sketch out the
> preliminary design?
No designs yet. A number of ideas kicking around: Drexler's fine motion
controller, Merkle's hydrocarbon metabolism, Zyvex's exponential
assembly,
general studies by von Neumann and Freitas, but nothing yet like a
detailed
complete design. Chris Phoenix had some interesting ideas about a
macroscale
self-replicator a few years back, made out of a liquid plastic that
would
be hardened by UV light.
There's nothing preventing anybody from thinking about a complete
design,
other than that it's a lot of work, and you'll need to make assumptions
that
may prove unwarranted when the enabling technologies fall into place. If
you're the first person to come up with a design, and it doesn't prove
to
be a hopeless mismatch when we get close enough to try building it,
you'll
probably have a pretty fair shot at defining what an assembler is
supposed
to look like.
> 2) Do we have a design yet for a Nanocomputer? What kind of instruction set
> would it have? How much memory? What programming languages will assemblers
> use?
There is currently a lot of work in molecular electronics going on.
There is
even hope of building useful molecular computers before we take off the
boxing gloves (following Merkle's analogy that pre-nano technology is to
atoms as a person wearing boxing gloves is to Lego blocks), thanks to
some
very ingenious developments by folks at Hewlett-Packard and UCLA. There
is
also interesting work in operating nanotubes as semiconductors, and
doping
them to form P-type and N-type materials, and building FETs
(field-effect
transistors) with them.
It's early to say too much about nanocomputers, but some things seem
pretty
clear. Memory should be cheap and plentiful. I'd guess that hundreds of
gigabytes of fast memory would be reasonable, with mass storage devices
offering much more. Processors will probably use very wide datapaths, so
the assembly languages might look like today's VLIW architectures. How
that
will influence high-level language design isn't clear to me.
> 3) What about networking? Trillions of assemblers are going to be working
> together to accomplish a goal. Will TCP/IP work?
IPv4 is limited to 32-bit addresses, so that won't scale very far. I
don't
remember the address size for IPv6. But we may not want to name or
number
individual assemblers, we may prefer to specify rules such that each
assembler figures out its own appropriate role based on its physical
position. Some years back there was a project along these lines at MIT
called "amorphous computing"; I don't know if it's still an active area
of
research. A few other organizations were doing similar things.
> 4) Finally, do we have computers powerful enough to simulate all of the
> above? Has anyone tried it yet?
As far as I know, the most powerful computers available are distributed
arrangements like SETI@Home, which work when the computing problem is
easily partitioned into small independent chunks. Lots of people run
background client processes on their home and office computers (such as
a
screensaver), each working on a chunk, and infrequently reporting back
to
a central server to deliver results and pick up a new chunk to work on.
Molecular modeling isn't a very good problem for this approach because
it's hard to break a big simulation into independent chunks; atoms tend
to influence one another over relatively large distances. There are
nevertheless one or two projects along these lines, such as the
Folding@Home project.
The next best thing is to run a big Linux cluster with a few hundred
processors (or however many you can afford). The largest things of this
type, I think, are at pharmaceutical companies, and I've seen one such
cluster which will soon be upgraded to about 300 processors. But it's
expensive to build such a cluster, and its time is therefore quite
valuable, and you'd have to make an awfully compelling sales pitch to
get the kind of time you'd need for this sort of simulation.
Christopher J. Phoenix
Very interesting questions. Below is mainly just my opinion; others
will have different answers.
"Rob K." wrote:
> 1) Do we have a design for an assembler? I realize that building one is
> still many years in the future, but is it too early to sketch out the
> preliminary design?
Merkle has done some work, both for a micro-scale sound-powered
proto-assembler and for a convergent-assembly factory. Start at
http://www.foresight.org/Conferences/MNT6/Papers/Merkle/ and look at the
bottom, where there are pointers to many Merkle assembler papers.
> 2) Do we have a design yet for a Nanocomputer? What kind of instruction set
> would it have? How much memory? What programming languages will assemblers
> use?
Drexler in _Nanosystems_ laid out the basic gates, registers, PLAs, and
other logic that are required to build a Von Neumann CPU. I don't think
there's much doubt we'll be able to make a design fairly easily when we
know what molecules/shapes the fabrication technology can build. In
other words, I think we can defer this problem for a few years without
hurting ourselves much.
> 3) What about networking? Trillions of assemblers are going to be working
> together to accomplish a goal. Will TCP/IP work?
No, the address space is only a few billion. But in a
convergent-assembly factory, you don't have trillions of assemblers.
You have trillions of workstations, all fastened down and very amenable
to hierarchical control. So there's no need to talk to all of them
directly. And there's no need even for networking protocols.
> 4) Finally, do we have computers powerful enough to simulate all of the
> above? Has anyone tried it yet?
I'm almost certain we don't have computers anywhere near powerful enough
to simulate at the atomic level. A microscale assembler has many
million atoms. Of course most of them will be structural, and so
simulatable without worrying about each bond. A mixed-mode simulator
might be possible. I don't know of anyone who's tried it yet. Merkle said
years ago that he thought it was time to start working on a simulator.
Chris
--
Chris Phoenix cpho...@best.com http://www.best.com/~cphoenix
Interests: nanotechnology, dyslexia, caving, filk, SF, patent
reform, stained glass, hang gliding
Fraser Orr
"Rob K." <rob...@yahoo.com> wrote in message
news:a51i3...@enews2.newsguy.com...
> 1) Do we have a design for an assembler? I realize that building one is
> still many years in the future, but is it too early to sketch out the
> preliminary design?
I have not heard of a complete design, though there are many
designs for the various pieces.
> 2) Do we have a design yet for a Nanocomputer? What kind of instruction
set
> would it have? How much memory? What programming languages will assemblers
> use?
The closest we have is Drexler's rod logic computer, and even
that is incomplete. However, in terms of the type of computer,
most likely it will behave much the same as the computers we have
today. How much memory? Probably depends on the application,
though most indications are that if you have reasonable molecular
level control, memory will be vastly capacious compared to today.
Programming languages? C++ and Java would be my guess. Nano
computers will likely be like silicon computers only smaller
faster and cheaper.
Sometimes people confuse esoteric computer architectures with
nanocomputers. For example Quantum computers or DNA computers.
Neither of these types of device is suitable for general computing
work, and probably never will be. Von Neumann was a very clever
guy indeed.
> 3) What about networking? Trillions of assemblers are going to be working
> together to accomplish a goal. Will TCP/IP work?
TCP/IP is a language of communication, it may work, however all
that goes with TCP/IP such as the network topologies we use today
will be entirely unsuitable for assemblers. There are simply too
many devices. Some entirely new way of programming will probably
have to be invented. Something that can heavily delegate down
a deep tree of control structures, perhaps. But more likely some
sort of system that allows a design to arise spontaneously without
a top down command structure. (Genetic algorithms being one such
possibility.)
In fact, now I think about it, TCP/IP probably won't work either
since quadrillions of devices would quickly outstrip the abilities
of such protocol to route messages effectively.
> 4) Finally, do we have computers powerful enough to simulate all of the
> above? Has anyone tried it yet?
Heck, we barely have computers powerful enough to calculate the
shape of a complex molecule like a protein, never mind a full
blown nanosystem!
Robert J. Bradbury
"Rob K." wrote:
> 1) Do we have a design for an assembler? I realize that building one is
> still many years in the future, but is it too early to sketch out the
> preliminary design?
No and perhaps. We need the molecular CAD software tools to assist in the
design. You could start designing it now. The time scale for Eric and Ralph
to design the ~2600 atom Fine Motion Controller was ~2 months using software
that is probably close to a decade old. You are probably dealing with 8
million
atoms, so that's about 6000 man-months (500 man years). You might be able
to knock off an order of magnitude (~600 man-months/50 man-years) if you
shoot for a design that has a significant amount of redundancy (e.g. once you
have designed one diamondoid sleeve bearing, designing one that is 20% larger
or 20% smaller doesn't take much time).
> 2) Do we have a design yet for a Nanocomputer? What kind of instruction set
> would it have? How much memory? What programming languages will assemblers
> use?
Not at the atomic scale, though the pieces have been described in greater
detail
than the assembler models. The first instruction sets would presumably be RISC
based since those produce smaller CPUs. Memory is always dictated by the
application, access time requirements, space available, etc. I expect that
most
nanomachine programming will initially be done in assembler.
> 3) What about networking? Trillions of assemblers are going to be working
> together to accomplish a goal. Will TCP/IP work?
If you can tolerate the CPU time and memory for the abstraction layer stack.
> 4) Finally, do we have computers powerful enough to simulate all of the
> above? Has anyone tried it yet?
I don't believe anyone has tried an 8 million atom simulation yet. I suspect
we
are roughly an order of magnitude below that, and even then the simulations
take quite some time on the fastest computers.
Robert
Toby Kelsey
Fraser Orr wrote:
> The closest we have is Drexler's rod logic computer, and even
> that is incomplete. However, in terms of the type of computer,
> most likely it will behave much the same as the computers we have
> today. How much memory? Probably depends on the application,
> Sometimes people confuse esoteric computer architectures with
> nanocomputers. For example Quantum computers or DNA computers.
> Neither of these types of device is suitable for general computing
> work, and probably never will be. Von Neumann was a very clever
> guy indeed.
True, but it seems likely that a nanocomputer will not use Von Neumann
architecture, which has clear bottlenecks. While the organisation of
construction will be as distributed as possible, any precalculated
grand design would probably best be performed by a single, relatively
large, computer.
Toby
[I asked for a clarification from Toby: he's saying that construction will use
local computers distributed over the construction area, while the "single,
relatively large computer" would use a parallelised architecture for multiple
threads of control to avoid the Von Neumann bottleneck. --CJP]
--
"They passed the open door of one large room in which rows of robots
were bent quietly over desks. Pelorat asked, 'What are they doing,
Bander?'
'Bookkeeping,' said Bander." - Isaac Asimov 1986 [Foundation and Earth]
Fraser Orr
"Toby Kelsey" <toby....@ntlworld.com> wrote in message
news:a5n7r...@enews4.newsguy.com...
> True, but it seems likely that a nanocomputer will not use Von Neumann
> architecture, which has clear bottlenecks. While the organisation of
> construction will be as distributed as possible, any precalculated
> grand design would probably best be performed by a single, relatively
> large, computer.
I doubt you are correct. Most likely nanocomputers will use
pretty much the same sorts of architectures we have today.
Highly parallel computer systems are simply regular computers
constructed in an array. It is true that the single threaded
processors we have had for fifty years have tended to result
in single threaded algorithms, and I agree that a whole new
slew of algorithms will be required to effectively use such
a resource. However, it is also true that not all computational
tasks are amenable to being calculated in parallel. And this
especially so if the interprocess communication time is non-
trivial (as would most likely be the case in massively parallel
nano-computational elements.) Perhaps and example might help
the non-computer programmers amongst us. It is probably true
that multiplying a pair of matrices is a highly parallel task,
however, since each individual task is small (a few multiplies
and additions) the cost of communicating the result back to
the result matrix store is usually vastly more expensive than
the calculation itself. Consequently, matrix multiplication is
simply never done in this fashion. It is true that you can
split the task up into larger chunks so that the overhead
is less, yet this can be an extremely complex process indeed.
There are some processes highly amenable to parallel calculation.
For example, ray tracing for image rendering. This is so because
the calculations and processes involved in each ray traced are
quite substantial, and consequently, the IPC overhead does not
completely outweigh the benefit of multiple processors.
So then it seems likely that most of the components of today's
computers (ALUs, CPUs, encoded instructions stored in memory,
common data memory, bus architectures etc.) will still be a major
part of the design of nanocomputers. In science you never say
never, but regular computer architectures seem fairly well suited
to application in a nanotech computer.
Christopher J. Phoenix
Fraser Orr wrote:
> .... It is true that the single threaded
> processors we have had for fifty years have tended to result
> in single threaded algorithms, and I agree that a whole new
> slew of algorithms will be required to effectively use such
> a resource. However, it is also true that not all computational
> tasks are amenable to being calculated in parallel. ....
Cooling is likely to require a certain limit on the density of any
particular form of Boolean computation, especially with 3D packing
available. Is it better to spread out your processing thinly over a
volume, or to concentrate it into nodes surrounded by coolant? I'm
betting on the latter--and a concentrated node sounds like a good fit
for a single-CPU Von Neumann architecture. Maybe a 1024-bit
architecture (splittable into 64 parallel 16-bit ALU ops), 100-stage
pipeline, 50-ALU CPU with integrated L1 cache (with 3D construction, the
cache could be right next to the registers) and up to 100 instruction
starts per clock cycle, but still a single CPU. (No, this is not
sarcasm.)
But what about reconfigurable logic? They've been talking about it for
years. Maybe with enough algorithm-churning they'll be able to make it
work. And what about VLIW? And don't forget virtual machines with JIT
compilers to native instructions. And wasn't there a processor that
could simulate other processors by preprocessing the instructions?
Think about the fact that, as implied above, single processors have been
trying harder and harder to be parallel, and have been flirting with
DSPs. Put it all together, add *really* cheap gates, and you get...
A whole mess of gates, that looks something like a bunch of ALUs and
registers (or maybe cache), that is reconfigurable on the fly, that can
stream computation through very fast--as long as there's a ton and a
half of pre-computation done. This is the "CPU". Feeding this over a
massive bus is a whole nest of logic designed to take an instruction and
data stream (or streams, from several programs/threads) and program the
CPU to do all the math in the right order. The nest of logic could
include all sorts of machine learning, compiling, reordering, and so
on. As such, it could be sequenced, and also distributed, and even
agoric with various methods bidding for use of the CPU to execute a
single chunk of code. The CPU is still monolithic, but it's only a
small part of the picture. And it's amazingly parallel and blindingly
fast; and the efficiency of use improves as the nest of logic upstream
acquires more modules.
It's true that parallel computing has not really caught on, because a
lot of algorithms are not inherently parallel. There's a human
bottleneck in programming them to be parallel (not to step on each
other's data). But you were also claiming a speed-of-light bottleneck
in the network to recombine the results. I believe it. There's also a
growing speed-of-light bottleneck between memory and CPU, and even
within the CPU itself, which is why we have to worry about all this in
the first place.
But meanwhile, standard CPUs have been getting more and more into
last-minute parallelism, and prediction/prefetching. And it's arguable
that the process of implementing last-minute parallelism has been
gate-limited and design-limited rather than speed-of-light limited like
memory or networks. How would you prefer to control the
hyper-meta-single-CPU described above? With intense amounts of
microcode? Complicated and specialized (brittle) VLIW compilers?
Perhaps it's best to use networked logic that can generate the CPU
scheduling at the last minute and send it over a network/bus to the CPU.
Does anyone have a figure for the percentage of silicon used to schedule
the operations in a modern CPU, and how that percentage has changed over
time? Does anyone know to what extent this logic is separable and
distributable?
Malcolm McMahon
On 1 Mar 2002 06:41:52 GMT, Toby Kelsey <toby....@ntlworld.com>
wrote:
>True, but it seems likely that a nanocomputer will not use Von Neumann
>architecture, which has clear bottlenecks.
People have been working on alternatives to the Von Neuman architecture
for a long time now; ideas such as data-flow architecture, but none has
been conspicuously successful. Most of the successful strategies have
employed large numbers of linked Von-Neumann machines or array
processors, which are simple Von Neuman machines with very wide data
busses.
Distributed processing, certainly, because it's best to do the
computation as close as possible to where the results are needed and
thus minimise the data transport overheads.
Fraser Orr
"Christopher J. Phoenix" <cpho...@best.com> wrote in message
news:a5pqf...@enews2.newsguy.com...
> Cooling is likely to require a certain limit on the density of any
> particular form of Boolean computation, especially with 3D packing
> available.
Cooling is a problem, but one that may largely be addressed by
reversible logic. However, the control that MNT gives over structure
could build an excellent honeycomb of coolant distribution channels.
One might imagine a coolant channel within a few nanometers of every
gate. Assuming that you could pump the stuff through fast enough
and provide a suitable heat sink then I suspect cooling will be
much less of a problem than it is today. (After all, all we do today
is glue a fan on top of the chip!)
Perhaps one might imagine perhaps 80% (SWAG) of the volume of future
processors being taken up by coolant channels.
> Is it better to spread out your processing thinly over a
> volume, or to concentrate it into nodes surrounded by coolant? I'm
> betting on the latter--and a concentrated node sounds like a good fit
> for a single-CPU Von Neumann architecture. Maybe a 1024-bit
> architecture (splittable into 64 parallel 16-bit ALU ops), 100-stage
> pipeline, 50-ALU CPU with integrated L1 cache (with 3D construction, the
> cache could be right next to the registers) and up to 100 instruction
> starts per clock cycle, but still a single CPU. (No, this is not
> sarcasm.)
I think that such an architecture would be of limited utility except
in special cases. Most calculations that humans deal with require
less than 32 bits, never mind 64, never mind 1024. Consequently, such
a processor would primarily run as 16 parallel 64 bit computers. As I
said before, parallelism of this kind is of limited utility. Further,
one could imagine a 100 stage pipeline, however it is unlikely to be
much use. Why? Because the likelihood of 100 independent instructions
in a row is very low, and consequently, the pipeline would constantly
stall when it had hardly any instructions in it.
> But what about reconfigurable logic? They've been talking about it for
> years. Maybe with enough algorithm-churning they'll be able to make it
> work.
One thing I think you are right about is this. One of the trends
in modern computing is to have a workhorse CPU at the center, but
to distribute entirely independent tasks off to helper CPUs. For
example, DMA devices copy data from such devices as disks and external
ports. It used to be that the data was copied by interrupting the
processor as each byte became available, and having the processor
copy it into memory. However, DMA systems take that load off the
processor. Similarly, graphics cards take the load of graphical
rendering off the processor, or DSPs the load of signal analysis.
Various other specialized devices like this have also been proposed and
built, and I suspect that this trend will continue, especially so
with the advent of hyper cheap manufacturing and the demise of Moore's
second law. However, in all these cases the extra processors were
designed for very specific tasks, the CPU is still required as the
general purpose processor. This takes me back to my original comment
about quantum computers and DNA computers. These are not general
purpose computers, they will never run, for example, Microsoft Word.
However, they fall into the category of the helper devices I described
above.
I can easily imagine a whole slew of useful devices that could be
plugged into the bus of a general purpose CPU. Your reconfigurable
logic device could be exceptionally useful in some very specialized
tasks, a neural net device could also be very useful, and much faster
and flexible than a software simulation. If memory serves me correctly
Big Blue beat Kasperov at chess, partly by having 256 specialized
devices for analyzing chess positions. However, at the end of the
day, they only cover a fraction of the applications that a general
purpose computer must deal with. It is surely a readily apparent
fact that the trend is toward general purpose software controlled
devices away from specialized hardware. (Goodbye, record players,
telephones, televisions and VCRs.)
> It's true that parallel computing has not really caught on, because a
> lot of algorithms are not inherently parallel.
I could be wrong Chris, but I think it is more fundamental than that,
that is to say, the problem is than many processes are inherently
serial, rather than the shortcoming being in our lack of algorithms.
(Though obviously some very important algorithms are very parallel,
such as the algorithm that is causing my fingers to type this stuff.)
> There's a human
> bottleneck in programming them to be parallel (not to step on each
> other's data). But you were also claiming a speed-of-light bottleneck
> in the network to recombine the results. I believe it. There's also a
> growing speed-of-light bottleneck between memory and CPU, and even
> within the CPU itself, which is why we have to worry about all this in
> the first place.
This could indeed be a problem, it takes light about 1e-18 seconds
to cross an atomic diameter so speeds above exaflops are going to
take some novel approaches. I don't know the answer to that.
> But meanwhile, standard CPUs have been getting more and more into
> last-minute parallelism, and prediction/prefetching. And it's arguable
> that the process of implementing last-minute parallelism has been
> gate-limited and design-limited rather than speed-of-light limited like
> memory or networks.
One possible approach that has been suggested is to run all
possible flow paths at once, and determine which to cull as
you carry on. For example, in C you might say:
i = function_that_takes_a_long_time();
switch (i)
{
case 1: f(1); break;
case 2: f(2); break;
case 3: f(3); break;
case 4: f(4); break;
}
In such a case you might run the "function_that_takes_a_long_time" on
one processor, and also, AT THE SAME TIME, run f(1), f(2), f(3) and
f(4) each on separate processors. When i was finally calculated, say
it is calculated as 1, then you simply kill off the processes calculating
f(2), f(3) and f(4) and throw away the work they did.
However, this can be very tricky to do in practice, since each
processor needs its own copy of the environment (such as the
memory, the external devices and so forth.) Also, obviously today
it is very costly in terms of requiring an exponentially growing
number of processors. (Exponential on how far forward you
pre-calculate.)
FWIW, a type of programming language, called a functional language,
which is designed in such a way that there is essentially no shared
context between different threads of control (that is to say it
produces no side-effects), may offer a way to implement the scheme
I just explained much more easily than the procedural languages
used for most software today. (If you are interested in this check
out www.haskell.org which is a purely functional language unlike
some earlier attempts like Lisp and ML, both of which offer global
variables as a shared context.)
> How would you prefer to control the
> hyper-meta-single-CPU described above? With intense amounts of
> microcode? Complicated and specialized (brittle) VLIW compilers?
> Perhaps it's best to use networked logic that can generate the CPU
> scheduling at the last minute and send it over a network/bus to the CPU.
Sounds like a major bottleneck to me!
> Does anyone have a figure for the percentage of silicon used to schedule
> the operations in a modern CPU, and how that percentage has changed over
> time? Does anyone know to what extent this logic is separable and
> distributable?
I don't know the percentage, but I'm sure it is very small, most of
a CPU is taken up with cache memory these days. A better question
might be does anybody know the percentage of design time dedicated
to designing the pipeline control mechanism. I'm sure it is quite
substantial.
Christopher J. Phoenix
Fraser Orr wrote:
>
> "Christopher J. Phoenix" <cpho...@best.com> wrote in message
> news:a5pqf...@enews2.newsguy.com...
> > Cooling is likely to require a certain limit on the density of any
> > particular form of Boolean computation, especially with 3D packing
> > available.
>
> Cooling is a problem, but one that may largely be addressed by
> reversible logic. However, the control that MNT gives over structure
> could build an excellent honeycomb of coolant distribution channels.
> One might imagine a coolant channel within a few nanometers of every
> gate. Assuming that you could pump the stuff through fast enough
> and provide a suitable heat sink then I suspect cooling will be
> much less of a problem than it is today. (After all, all we do today
> is glue a fan on top of the chip!)
>From Nanosystems 12.8: "In nanomechanical systems of sufficiently large
scale, the ~10^12 W/m^3 power dissipation density of ~1 GHz
nanomechanical logic systems described here exceeds any possible means
of cooling." He suggests that convective cooling can cool (10^5 W) a 1
cm^3 volume containing 10^12 CPU-scale systems. Note that less than 1/8
of the volume is taken up by the 400-nm systems. That implies that
interprocessor signals are propagating twice as far as they would
without cooling channels.
Granted, we have no clue how to use 10^12 CPUs for most problems. And
even a mega-CPU such as I described would not even come close to needing
internal cooling. But convective cooling does not scale; if you need
more than 10^21 MIPS in a small volume you may be out of luck.
> Perhaps one might imagine perhaps 80% (SWAG) of the volume of future
> processors being taken up by coolant channels.
Yep. But remember that this substantially increases speed-of-light lag.
> > .... a concentrated node sounds .... Maybe a 1024-bit ....
>
> I think that such an architecture would be of limited utility except
> in special cases. Most calculations that humans deal with require
> less than 32 bits, never mind 64, never mind 1024.
Why then are we already building 64-bit processors? We'll certainly
want more than 64 bits of addressable memory.
> Consequently, such
> a processor would primarily run as 16 parallel 64 bit computers. As I
> said before, parallelism of this kind is of limited utility.
Sounds perfect for graphics and other sensory processing. That will be
one of the main applications.
> Further,
> one could imagine a 100 stage pipeline, however it is unlikely to be
> much use. Why? Because the likelihood of 100 independent instructions
> in a row is very low, and consequently, the pipeline would constantly
> stall when it had hardly any instructions in it.
Depends how many processes you're running. And there are tricks like
result forwarding that can allow dependent instructions to go through
the pipeline. But I admit I got overconfident here. I'll say 20, or
else asynchronous (self-clocked) logic.
> > But what about reconfigurable logic? They've been talking about it for
> > years. Maybe with enough algorithm-churning they'll be able to make it
> > work.
>
> One thing I think you are right about is this. One of the trends
> in modern computing is to have a workhorse CPU at the center, but
> to distribute entirely independent tasks off to helper CPUs.
I believe mainframes used to have an I/O processor separate from the CPU
that offloaded most of the memory-transfer tasks. My impression is that
this was a very powerful chunk of hardware, far beyond a simple DMA
controller. For example, the I/O processor would handle all of the
terminals. The CPU would run in batch mode.
> graphics cards ... DSPs ... Various other specialized devices ...
> However, in all these cases the extra processors were
> designed for very specific tasks, the CPU is still required as the
> general purpose processor. This takes me back to my original comment
> about quantum computers and DNA computers. These are not general
> purpose computers, they will never run, for example, Microsoft Word.
> However, they fall into the category of the helper devices I described
> above.
I agree about quantum and DNA computers. (I'm still not convinced that
DNA computers are good for much. They're certainly not good for
real-world 2^N problems, despite much hype to that effect.)
> I can easily imagine a whole slew of useful devices that could be
> plugged into the bus of a general purpose CPU. Your reconfigurable
> logic device could be exceptionally useful in some very specialized
> tasks, a neural net device could also be very useful, and much faster
> and flexible than a software simulation. If memory serves me correctly
> Big Blue beat Kasperov at chess, partly by having 256 specialized
> devices for analyzing chess positions. However, at the end of the
> day, they only cover a fraction of the applications that a general
> purpose computer must deal with. It is surely a readily apparent
> fact that the trend is toward general purpose software controlled
> devices away from specialized hardware. (Goodbye, record players,
> telephones, televisions and VCRs.)
Yep. Also, at least in low-end systems, the CPU is quite usable to
simulate things we can do perfectly well in external hardware.
> > It's true that parallel computing has not really caught on, because a
> > lot of algorithms are not inherently parallel.
>
> I could be wrong Chris, but I think it is more fundamental than that,
> that is to say, the problem is than many processes are inherently
> serial, rather than the shortcoming being in our lack of algorithms.
I think that's what I meant to say.
> This could indeed be a problem, it takes light about 1e-18 seconds
> to cross an atomic diameter so speeds above exaflops are going to
> take some novel approaches. I don't know the answer to that.
If you use bosons, you can use the same computing elements for many
computations at the same time, at slightly different frequencies. Note
this is *not* quantum computing: it's not superposition of states,
simply addition of states. At sufficiently low temperatures, some
nuclei can be bosons, allowing shorter wavelengths than photons.
(Warning: this is speculative and not well grounded.)
> One possible approach that has been suggested is to run all
> possible flow paths at once, and determine which to cull as
> you carry on. ....
> However, this can be very tricky to do in practice, ....
Yep, hence the mass of logic feeding the CPU.
> > Perhaps it's best to use networked logic that can generate the CPU
> > scheduling at the last minute and send it over a network/bus to the CPU.
>
> Sounds like a major bottleneck to me!
Why? The logic can be massively parallel, and the bus can be quite
wide. Of course it'd have to be asynchronous or self-clocking.
> > Does anyone have a figure for the percentage of silicon used to schedule
> > the operations in a modern CPU, ....
>
> I don't know the percentage, but I'm sure it is very small, most of
> a CPU is taken up with cache memory these days.
They're putting cache in the CPU? I knew they were putting L1 in the
same package, but still on a separate chip.
> A better question
> might be does anybody know the percentage of design time dedicated
> to designing the pipeline control mechanism. I'm sure it is quite
> substantial.
Yah... I wasn't suggesting that humans design the feed-in logic.
Fraser Orr
> I agree about quantum and DNA computers. (I'm still not convinced that
> DNA computers are good for much. They're certainly not good for
> real-world 2^N problems, despite much hype to that effect.)
Chris, I recently read the article below, and it reminded me of
your comment above. The article is a description of the use of
a DNA computer to solve an NP complete problem. Although it seems
to be a rather non-useful problem in itself, there is a proof
out there somewhere that says solving one NP complete problem
allows you to solve any NP-complete problem. The ability to solve
some such problems (particularly the so called traveling salesman
problem) would be an extremely valuable thing. Probably it would
be worth billions of dollars. In fact it might even greatly help
the development of nanotech itself, since protein folding is also
an NP-complete problem.
Of course 20 variables is well within the abilities of even small
computers, but the point is that in a DNA computer the time to
solution scales linearly, rather than exponentially as it does in
a conventional computer.
FWIW. <http://physicsweb.org/article/news/6/3/11>
erincss
>Fraser Orr
>The ability to solve
>some such problems (particularly the so called traveling salesman
>problem) would be an extremely valuable thing. Probably it would
>be worth billions of dollars. In fact it might even greatly help
>the development of nanotech itself, since protein folding is also
>an NP-complete problem.
>
>Of course 20 variables is well within the abilities of even small
>computers, but the point is that in a DNA computer the time to
>solution scales linearly, rather than exponentially as it does in
>a conventional computer.
>
>FWIW. <http://physicsweb.org/article/news/6/3/11
>
Let's say the development of DNA computers allows us to actually design
proteins that fold up into molecular scale machine parts and objects, to build
a first generation assembler. Depending on what variables, how long, in your
estimation, would it be, to get from there to actual assemblers made of harder
stuff, like diamondoid?
Eugene Leitl
On 18 Mar 2002, Fraser Orr wrote:
> Of course 20 variables is well within the abilities of even small
> computers, but the point is that in a DNA computer the time to
> solution scales linearly, rather than exponentially as it does in
> a conventional computer.
That's not true. Unlike entangled qubits we're talking about base-base
interaction in a solvated linear biopolymer. The parallelism is all in the
intermolecular interaction, and of course the batch size is limited,
unless you have a lot of time to burn. You'd do much, much better by using
that atoms to build a few moles of molecular circuitry, which is not
subject to viscous drag, being electronic in nature.
It's not a solution to NP, it's brute-forcing it. And of course it doesn't
scale, it's the nature of NP.
[I agree completely. Note what the article said: "Each of the possible
1,048,576 solutions were then represented by much longer strands of specially
encoded DNA". What happens when you have 1,307,674,386,000 possible solutions?
That's the Travelling Salesman problem with only 15 cities. Around 24 cities,
it requires *kilograms* of DNA, each strand individually synthesized. This
type of DNA computer is utterly worthless for any non-toy exponential problem
requiring a precise answer. If the problem does not require a precise answer,
the computer has to compete with genetic algorithms and clever math tricks that
can be implemented just fine on silicon. I'm surprised these people still have
funding--why hasn't anyone noticed that their method can't possibly scale to a
useful size? --CJP]
Steve Lenhert
Eugene Leitl wrote:
> It's not a solution to NP, it's brute-forcing it. And of course it doesn't
> scale, it's the nature of NP.
Well put Eugene. Its also interesting to note how much "brut force" there
is at the bottom of a test-tube. Immagine if one were to use even more
brut-force, and make use of, for instance algorithmic self-assembly
http://seemanlab4.chem.nyu.edu/XOR.html
or perhaps the one-dimensional electronic properties of DNA
http://www.nanoword.net/library/weekly/aa030100a.htm
> [I agree completely. Note what the article said: "Each of the possible
> 1,048,576 solutions were then represented by much longer strands of specially
> encoded DNA". What happens when you have 1,307,674,386,000 possible
> solutions? That's the Travelling Salesman problem with only 15 cities.
> Around 24 cities, it requires *kilograms* of DNA, each strand individually
> synthesized. This type of DNA computer is utterly worthless for any non-toy
> exponential problem requiring a precise answer. If the problem does not
> require a precise answer, the computer has to compete with genetic algorithms
> and clever math tricks that can be implemented just fine on silicon. I'm
> surprised these people still have funding--why hasn't anyone noticed that
> their method can't possibly scale to a useful size? --CJP]
To answer the co-moderators question about "scaling" and "funding".
First, let us distingush between science and technology:
Science:
The pursuit of knowledge and understanding, from the Latin term
'scientia', which means 'knowledge'.
http://www.nanoword.net/library/def/science.htm
Technology:
The application of knowledge and understanding in order to control
processes and fabricate products, from the Greek root
'Techni', meaning 'art'.
http://www.nanoword.net/library/def/technology.htm
Notice that studies of these DNA increase our knowledge of a
variety of fields, and
therefore they continue to be accepted into scientific journals. e.g.
http://www.nature.com/cgi-taf/DynaPage.taf?file=/nature/journal/v414/n6862/abs/414430a0_fs.html
Once that knowledge is gained, tangible products naturally follow, e.g.
http://www.pcworld.com/news/article/0,aid,82403,00.asp
(I know its not exactly what we expect from a "computer", but at least
its not science fiction!)
Now, let us note that Zyvex (the MNT company) defines MNT as
"the science of manipulating molecules and atoms to create precise
structures." Assuming then that MNT is a "science" (don't ask me
why they don't call it molecular nano-science), I searched the
interdisciplinary, field defining scientific journals known as
'Science' and 'Nature' for use of the term "MNT" and found... nothing
- not counting the article about the Moist Non-acidic Tundra (MNT).
Now, I have a frequently unanswered question:
Why is Zyvex funded?
Steve Lenhert
http://www.nanoword.net
Defining Nanotechnology
erincss
>Steve Lenhert
>Now, I have a frequently unanswered question:
>
>Why is Zyvex funded?
>
>Steve Lenhert
>http://www.nanoword.net
>Defining Nanotechnology
Most likely because Zyvex's plan seems to be more straight forward, and
definitive. Ie mems systems until you reach the nanoscale, then nanoscale
systems bottom up, while at the same time, working on the building blocks for
bottom up mnt.
Fraser Orr
"Steve Lenhert" <st...@cmp-cientifica.com> wrote in message
news:a7abu...@enews2.newsguy.com...
> Now, I have a frequently unanswered question:
>
> Why is Zyvex funded?
The answer is simple. It is funded by an immensely rich guy
who doesn't want to die, and probably figures that he can't
take his big pile of money with him.
Gina Miller
Nanogirl News
(nano + plus)
March 22, 2002
Nanocrystals' trapped atoms technology studied at Oak Ridge Labs. A
whole new technology awaits exploration with the discovery of a
technique for trapping single atoms, say scientists at the Department
of Energy's Oak Ridge National Laboratory (ORNL). ORNL's Mike Barnes,
Thomas Thundat and Adosh Mehta have collaborated with Ramesh Bhargava
of Nanocrystals Technology in Briarcliff, N.Y., to cage single
europium atoms in nanocrystals not much larger than the atoms
themselves.
(Nanotech-Planet 3/21/03)
http://www.nanotech-planet.com/features/article/0,4028,6571_996291,00.html
Nano-Tex Markets Brand To Become "Intel Inside" OF NANO MATERIALS.
Nano-Tex LLC has an unusual position in the nanomaterials industry.
It's not just its products, polymers engineered to attach to fibers
to make clothes stain-proof, wrinkle-free, absorbent or more
comfortable. Nor is it its worldwide business partnerships with
textile mills and fabric manufacturers. "We're trying to position
ourselves to be an ingredient brand, like Intel," said George
Henderson III, chief executive of Nano-Tex and the textile company
Burlington Industries, both based in Greensboro, N.C. Burlington owns
51 percent of Nano-Tex. (Small Times 3/23/02)
http://www.smalltimes.com/document_display.cfm?document_id=3341
Magnetism to its lowest terms. While an ideal infinite chain of atoms
cannot sustain ferromagnetic order at nonzero temperature, an
international team of physicists shows that the finite length of the
chains and magnetic anisotropy barriers stabilize ferromagnetism at
finite temperatures in monatomic cobalt chains. The persistence of
ferromagnetism in monatomic chains has important consequences for the
design and properties of magnetic nanostructures, a very active field
of physical research and development today. (Max Planck Society
3/21/02)
http://www.mpg.de/news02/news0206.htm
British Law Would Make Organ Donation Automatic Upon Death. People
should have their organs automatically removed for transplants after
they die unless they had expressed objections when alive, according
to a Bill to go before the Commons today. The 10-minute rule Bill,
supported by doctors' leaders, reaches MPs amid fears that people are
dying needlessly because of a shortage of donors. (Rense 3/21/02)
http://www.rense.com/general21/britishlawwould.htm
Mainstream Business, Government Interest Is Growing In Nano Research
And Investing. The next big thing to come out of this birthplace of
high tech could be small: Think tiny molecular delivery devices for
medicines. Or "smart" dust that can monitor people without being
detected. Maybe supercomputers the size of grains of salt. The
mind-bending ideas seem straight out of works of science fiction, but
some out here think they may be possible in the near future as
interest grows in something called nanotechnology. (Small Times
3/21/02)
http://www.smalltimes.com/document_display.cfm?document_id=3338
or see:
http://www.newsbytes.com/news/02/175376.html
Bridgestone develops nanotech display. Japan's automobile tire maker
Bridgestone Corp. has developed a material that can be used to make
electric display panels consuming one-five hundredth the power of
conventional LCDs (liquid crystal displays), the company said
Wednesday.
The new material, developed using nanotechnology, could be used to
make display panels which respond to moving images 100 times faster
than LCDs, for around half the cost, a Bridgestone spokesman said.
(IDG 3/20/02)
http://www.idg.net/ic_833294_1773_1-3921.html
Korea budgets $1.56 billion for nano R&D. South Korea's Ministry of
Science and Technology said it plans to invest about $1.56 billion in
2002 in nanotechnology research and development facilities. South
Korea's Ministry of Science and Technology said it plans to invest
about $1.56 billion in 2002 in nanotechnology research and
development facilities. The funds will also be used to train
engineers in the emerging field, and to support specific
nanotechnology research projects. (EE Times 3/20/02)
http://www.eet.com/at/news/OEG20020320S0041
Scientist wires his nervous system to computer. A British robotics
scientist has wired up his nervous system to a computer in an
experiment he hopes will eventually give paralyzed people more
control over their own bodies. Kevin Warwick, professor of
cybernetics at Reading University in southern England, has had minute
sensors implanted in the main nerve in his left arm and hooked up to
a radio transceiver that will send and receive messages from a
computer.
"We have a serious goal, a very medical goal of helping people with
spinal injuries and the like where there is a break in the nervous
system," Warwick told Reuters in a telephone interview.
"We will be sending signals from the nervous system to the computer
by radio and back the other way." (HoustonChronical 3/22/02)
http://www.chron.com/cs/CDA/story.hts/tech/news/1303782
Cool Carbon Foam New Material Leads to Cool Ideas. Feeling hot? One
day, a simple-looking block of solid foam could be the key to cooling
you - or anything else - down. Rrecently, researchers at the Oak
Ridge National Labs (ORNL) in Tennessee have developed a version of
carbon foam that acts as a super-conductor rather than an insulator
of heat. (ABCNews 3/22/02)
http://abcnews.go.com/sections/scitech/CuttingEdge/cuttingedge020322.html
Self-configuring array enables atomic-scale fabrication. A
next-generation configurable circuit architecture is being proposed
by a group of designers at startup Cell Matrix Corp. The company is
building small prototype chips based on the concept. The architecture
goes beyond basic FPGAs by building arrays of "cells" rather than
simply reconfigurable gates. Each cell has a small amount of logic
and local memory and communicates with its nearest neighbor. (EETimes
3/19/02)
http://www.eet.com/at/news/OEG20020319S0029
Lord of the Robots. The director of MIT's Artificial Intelligence Lab
says the age of smart, mobile machines is already beginning. You just
have to know where to find them-say, in oil wells. In the
not-too-distant future, a lot more people may be living with
technologies that Brooks's lab is developing. To help make pervasive
computing a reality, researchers in his lab and MIT's Laboratory for
Computer Science are developing-in an effort Brooks codirects called
Project Oxygen-the requisite embeddable and wearable devices,
interfaces and communications protocols. (TechReview 4/02)
http://www.techreview.com/articles/qa0402.asp
Lab-grown fish chunks could feed astronauts. Cooking up ways to feed
explorers on long trips in space, scientists have coaxed chunks of
fish meat to grow in the laboratory. The technique could lead to the
production of copious amounts of protein for consumption without the
messy and involved business of killing fish or livestock. "This could
save you having to slaughter animals for food," said Morris
Bejaminson, a bioengineer and the leader of the NASA-funded project.
(CNN 3/22/02) http://www.cnn.com/2002/TECH/space/03/22/fish.food/index.html
New lab is first at MIT to test tiny devices. A new NanoMechanical
Technology Laboratory at MIT will allow Institute researchers to
probe the mechanical properties of surfaces and devices at the atomic
and molecular scale. The laboratory "will have unique capabilities
for studying the properties of the tiny world," said Subra Suresh,
head of the Department of Materials Science and Engineering (DMSE).
For example, it will be home to the Institute's first nanoindenters,
machines that probe and measure the properties of surfaces of
engineering and biological materials. (MIT 3/15/02)
http://web.mit.edu/newsoffice/nr/2002/nanolab.html
Present at the Creation. A Russian-Livermore collaboration has added
two new elements,
114 and 116, to the periodic table. Both have the comparatively long
lifetimes predicted by long-held superheavy element theory. The
Livermore researchers are continuing its work to explore the
southwest shores of the island of stability. With funding from the
Laboratory Directed Research and Development program, they have begun
efforts to add elements 115 and 113 to the periodic table. They are
in the process of sending 22 milligrams of pure americium-243 to
Dubna for the work on element 115. (Science and Technology Review
3/8/02)
http://www.llnl.gov/str/JanFeb02/Moody.html
A New Form of Matter. Scientists have created a new kind of matter:
It comes in waves and bridges the gap between the everyday world of
humans and the micro-domain of quantum physics.
It's not often that you get to be around for the birth of a new kind
of matter, but when you do, the excitement is tremendous. "To see
something which nobody else has seen before is thrilling and deeply
satisfying. Those are the moments when you want to be a scientist,"
says Wolfgang Ketterle, a physicist at MIT and one of the first
scientists to create a new kind of matter called Bose-Einstein
condensates. (NASA 3/20/02)
http://science.nasa.gov/headlines/y2002/20mar_newmatter.htm?list154233
Helium Nanoclusters Retain Full Magnetic Qualities. University of
Florida scientists report discovering a curious anomaly of magnetism
in an article that appears in the current edition of Physical Review
Letters. While "Novel Magnetism in He3 nano-Clusters" is a textbook
example of basic science, it could have implications for
nanotechnology, the fast-evolving science of creating ultra small
devices for biomedical and other applications, said UF physics
Professor Dwight Adams. (UniSci 3/20/02)
http://unisci.com/stories/20011/0320016.htm
(Conference) April 29 - May 1, 2002 National Nanotechnology
Initiative: From Vision to Commercialization. is designed to provide
a comprehensive and in-depth look at the NNI. The focus of this
exciting conference is to look at what areas are being funded and
why, what the results of research have been to-date, and what avenues
appear to hold commercial promise.
http://www.infocastinc.com/Nanotech/home.htm
Nanophase Technologies Corporation (Nasdaq: NANX), a leader in
nanomaterials and nanoengineered products, today provided additional
guidance on expected revenue for the first and second quarter of
2002. The company now expects revenues of approximately $1.4 million
for the first quarter and $1.6 million in the second quarter, or
approximate revenue of $3 million during the first half of 2002.
Planned revenue growth in the first half of 2002 represents an
expected increase of 40% compared to the same period of 2001.
(Nanophase 3/13/02)
http://www.nanophase.com/new_2002_mar_13.shtml
Researchers discover new mechanism that targets and destroys abnormal RNA
HHMI researchers have identified a new mechanism that cells use to
recognize and destroy messenger RNA (mRNA)that contains errors.
Although this is a normal RNA-surveillance mechanism, the researchers
suggest that it might interfere with some drug treatments for cystic
fibrosis and other genetic diseases. (Howard Hughes Medical Institute
3/22/02)
http://www.hhmi.org/news/dietzparker.html
Science Close To Viewing The Beginning Of Time, University Of
Washington Cosmologist Says
When it comes to inflation, cosmologists are pondering a future that
probably would leave even Alan Greenspan scratching his head. Of
course, the Federal Reserve chairman is merely concerned with
economic policy and hasn't had to stare down the complexities of how
the universe pumped up after the Big Bang. (University of Washington
3/21/02)
http://www.washington.edu/newsroom/news/2002archive/03-02archive/k032102a.ht
ml
X-ray microscope can image crystalline grains in three dimensions. A
lensless X-ray microscope that can create three-dimensional images of
micron-size samples has been developed by scientists at the
University of Illinois. The instrument can be used in metallurgical
and semiconductor applications, and for studying the early growth
stages of protein crystals. (University of Illinios at
Urbana-champaign 3/18/02)
http://www.news.uiuc.edu/scitips/02/0318xray.html
UCLA Molecular Biologists Unravel Mysteries of 'Factory of Life'. In
their Nature paper, Lake and Simonson explain the molecular details
of elongation, including the location and movement of more than
10,000 atoms. In addition, they have located a novel binding site for
transfer RNA (tRNA) when it enters the ribosome. With the new
knowledge, it may become possible to make modifications in parts of
the translation process to suppress lethal mutations and design new
proteins to counteract the defects that cause numerous diseases, Lake
said. (UCLA news 3/21/02)
http://www.uclanews.ucla.edu/Docs/3097.htm
Gina "Nanogirl" Miller
Nanotechnology Industries
http://www.nanoindustries.com
Personal: http://www.nanogirl.com
Foresight Senior Associate
nano...@halcyon.com
"Nanotechnology: Solutions for the future."