Re: [math-fun] reversible computation; was 32-bit computers almost gone
I haven't followed the "reversible" research for several years now, but I think that this research somehow they got off on the wrong track to begin with, and never quite got back. In the '90's, the goal was to *lower* the energy dissipation, but not eliminate it (for reversible circuits). Thus, they did really slow logic, and some amount of charge recovery. But -- as you point out -- if you're using a highly non-linear device like a diode -- there are severe limits on what can be achieved. There needs to be some amount of non-linearity/error-correction in order to act as a refrigerator to reject the "heat" (in the form of noise) that intrudes into the computation. So we run into typical Maxwell's Demon issues. So we need to move to computational elements which are more isolated from the environment to begin with -- i.e., these computations are more "insulated" from the environment. It so happens that the requirements for reversible computation overlap those for quantum computations almost completely. After all, quantum computations are reversible, so they (using non-quantum modes) can emulate classical reversible computations. The real question is: if you have to do many of the same things for reversible computations as for quantum computations, then why not go directly to quantum computations -- which are strictly more powerful? That's probably why most non-quantum reversible computation research stopped a number of years ago. Now one aspect of reversible computation research hasn't (to my knowledge) been worked on very much: the emulation of complex reversible circuits using compositions of simpler reversible elements. Sooner or later, such compositions will be required for quantum computations, as well, but I guess much of this research has been tabled until actual quantum computers can be built. This is a pity, as it would be interesting to better understand the *synthesis* of complex reversible circuits from elementary reversible gates. Such an understanding may provide more insight into why certain reversible circuits are *hard to invert* -- e.g., encryption circuits. At 10:24 AM 7/6/2016, Richard Howard wrote:
1. Cheap means something different if you are talking about ubiquitous IOT nodes.
Large amount of embedded memory are very expensive and external RAM chips really break the bank for cost/size etc.
Trend is to single chip solutions with MCU/radio/memory. Not compatible with large quantities of DRAM.
As to delta in linewidth--the news is that the end is near because we are running out of atoms. Check the ITRS roadmap and Intel.
I have a nice photo of a CMOS device with 142 atoms from source to drain and 6 atoms of oxide (you can see the dots).
That was 20 years ago in research--we are there on-schedule in production.
2. Sounds wonderful, but for IOT you need a processor with a very large production run to get the cost down. Maybe someday SIMD single chip solutions will be there, but not in the near term.
3) Reversible computers are delightful ideas, but there are fundamental issues running them at room temperature.
Because of the kT/q slopes in diodes, you can't actually recover the energy. Hardware was made 20 years ago and shows that nicely. Now at mK temperatures, that is a different story.
--R
On Wed, Jul 6, 2016 at 11:48 AM Henry Baker <hbaker1@pipeline.com> wrote:
Yes, but...
1. Memory is cheap. Even uP will have decent sized memories. Every delta in line widths produces at least O(delta^2) in memory size, if not O(delta^3).
2. SIMD-type processors are cheap and quite energy-efficient. There's really no reason for a traditional 16-bit processor anymore.
3. Most crypto codes are 1-1 functions, and therefore *reversible*. In theory, reversible functions can be computed with asymptotically zero energy dissipation.
At 07:31 AM 7/6/2016, Richard Howard wrote:
BTW, there is an interesting mathy problem here--how do you do security when you have only a few microjoules and a few kB of memory in a 16 bit processor?
RSA is not even remotely feasible.
If you have superconducting interconnect, then good inductors are possible. If you can make good inductors, you can make high speed reversible circuits easily (you are not limited to adiabatic processing). There still is a huge advantage to being able to have small amounts of irreversible logic (that is, to throw away some information). I think of it in the same vein as functional programming. You want to build programs in a functional style except for a very few places, where it really really pays off to have side effects. I have been pushing DARPA to get serious about superconducting on-chip interconnect for about 20 years. In addition to good inductors (and hence high speed reversible circuits) it also speeds up on-chip communication from about 0.03c to about 0.6c. My more cynical view is that there is already such a program, but that it is classified. Along with nano-vacuum tubes.
On Jul 6, 2016, at 2:43 PM, Henry Baker <hbaker1@pipeline.com> wrote:
I haven't followed the "reversible" research for several years now, but I think that this research somehow they got off on the wrong track to begin with, and never quite got back.
In the '90's, the goal was to *lower* the energy dissipation, but not eliminate it (for reversible circuits).
Thus, they did really slow logic, and some amount of charge recovery.
But -- as you point out -- if you're using a highly non-linear device like a diode -- there are severe limits on what can be achieved.
There needs to be some amount of non-linearity/error-correction in order to act as a refrigerator to reject the "heat" (in the form of noise) that intrudes into the computation. So we run into typical Maxwell's Demon issues.
So we need to move to computational elements which are more isolated from the environment to begin with -- i.e., these computations are more "insulated" from the environment.
It so happens that the requirements for reversible computation overlap those for quantum computations almost completely. After all, quantum computations are reversible, so they (using non-quantum modes) can emulate classical reversible computations.
The real question is: if you have to do many of the same things for reversible computations as for quantum computations, then why not go directly to quantum computations -- which are strictly more powerful?
That's probably why most non-quantum reversible computation research stopped a number of years ago.
Now one aspect of reversible computation research hasn't (to my knowledge) been worked on very much: the emulation of complex reversible circuits using compositions of simpler reversible elements.
Sooner or later, such compositions will be required for quantum computations, as well, but I guess much of this research has been tabled until actual quantum computers can be built. This is a pity, as it would be interesting to better understand the *synthesis* of complex reversible circuits from elementary reversible gates. Such an understanding may provide more insight into why certain reversible circuits are *hard to invert* -- e.g., encryption circuits.
At 10:24 AM 7/6/2016, Richard Howard wrote:
1. Cheap means something different if you are talking about ubiquitous IOT nodes.
Large amount of embedded memory are very expensive and external RAM chips really break the bank for cost/size etc.
Trend is to single chip solutions with MCU/radio/memory. Not compatible with large quantities of DRAM.
As to delta in linewidth--the news is that the end is near because we are running out of atoms. Check the ITRS roadmap and Intel.
I have a nice photo of a CMOS device with 142 atoms from source to drain and 6 atoms of oxide (you can see the dots).
That was 20 years ago in research--we are there on-schedule in production.
2. Sounds wonderful, but for IOT you need a processor with a very large production run to get the cost down. Maybe someday SIMD single chip solutions will be there, but not in the near term.
3) Reversible computers are delightful ideas, but there are fundamental issues running them at room temperature.
Because of the kT/q slopes in diodes, you can't actually recover the energy. Hardware was made 20 years ago and shows that nicely. Now at mK temperatures, that is a different story.
--R
On Wed, Jul 6, 2016 at 11:48 AM Henry Baker <hbaker1@pipeline.com> wrote:
Yes, but...
1. Memory is cheap. Even uP will have decent sized memories. Every delta in line widths produces at least O(delta^2) in memory size, if not O(delta^3).
2. SIMD-type processors are cheap and quite energy-efficient. There's really no reason for a traditional 16-bit processor anymore.
3. Most crypto codes are 1-1 functions, and therefore *reversible*. In theory, reversible functions can be computed with asymptotically zero energy dissipation.
At 07:31 AM 7/6/2016, Richard Howard wrote:
BTW, there is an interesting mathy problem here--how do you do security when you have only a few microjoules and a few kB of memory in a 16 bit processor?
RSA is not even remotely feasible.
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
Superconducting have some serious limitations, of which temperature is only one. They have significant ac loss from flux motion and the long penetration depths associated with high temperature materials means that they are not good at confining magnetic fields--as is necessary for small inductors and transmission lines. Optical interconnects would be good alternatives, if low energy photons weren't so big and coupling strengths so small. Quantum computation sounds wonderful (in the original sense of "wonder"), but unlikely to be useful in leaf nodes of the IOT in my lifetime. Perhaps we can use molecular-based QC using biological quantum effects, like seen in chlorophyll and grow the devices. --R On Wed, Jul 6, 2016 at 5:42 PM Tom Knight <tk@mit.edu> wrote:
If you have superconducting interconnect, then good inductors are possible. If you can make good inductors, you can make high speed reversible circuits easily (you are not limited to adiabatic processing). There still is a huge advantage to being able to have small amounts of irreversible logic (that is, to throw away some information). I think of it in the same vein as functional programming. You want to build programs in a functional style except for a very few places, where it really really pays off to have side effects.
I have been pushing DARPA to get serious about superconducting on-chip interconnect for about 20 years. In addition to good inductors (and hence high speed reversible circuits) it also speeds up on-chip communication from about 0.03c to about 0.6c. My more cynical view is that there is already such a program, but that it is classified. Along with nano-vacuum tubes.
On Jul 6, 2016, at 2:43 PM, Henry Baker <hbaker1@pipeline.com> wrote:
I haven't followed the "reversible" research for several years now, but I think that this research somehow they got off on the wrong track to begin with, and never quite got back.
In the '90's, the goal was to *lower* the energy dissipation, but not eliminate it (for reversible circuits).
Thus, they did really slow logic, and some amount of charge recovery.
But -- as you point out -- if you're using a highly non-linear device like a diode -- there are severe limits on what can be achieved.
There needs to be some amount of non-linearity/error-correction in order to act as a refrigerator to reject the "heat" (in the form of noise) that intrudes into the computation. So we run into typical Maxwell's Demon issues.
So we need to move to computational elements which are more isolated from the environment to begin with -- i.e., these computations are more "insulated" from the environment.
It so happens that the requirements for reversible computation overlap those for quantum computations almost completely. After all, quantum computations are reversible, so they (using non-quantum modes) can emulate classical reversible computations.
The real question is: if you have to do many of the same things for reversible computations as for quantum computations, then why not go directly to quantum computations -- which are strictly more powerful?
That's probably why most non-quantum reversible computation research stopped a number of years ago.
Now one aspect of reversible computation research hasn't (to my knowledge) been worked on very much: the emulation of complex reversible circuits using compositions of simpler reversible elements.
Sooner or later, such compositions will be required for quantum computations, as well, but I guess much of this research has been tabled until actual quantum computers can be built. This is a pity, as it would be interesting to better understand the *synthesis* of complex reversible circuits from elementary reversible gates. Such an understanding may provide more insight into why certain reversible circuits are *hard to invert* -- e.g., encryption circuits.
At 10:24 AM 7/6/2016, Richard Howard wrote:
1. Cheap means something different if you are talking about ubiquitous IOT nodes.
Large amount of embedded memory are very expensive and external RAM chips really break the bank for cost/size etc.
Trend is to single chip solutions with MCU/radio/memory. Not compatible with large quantities of DRAM.
As to delta in linewidth--the news is that the end is near because we are running out of atoms. Check the ITRS roadmap and Intel.
I have a nice photo of a CMOS device with 142 atoms from source to drain and 6 atoms of oxide (you can see the dots).
That was 20 years ago in research--we are there on-schedule in production.
2. Sounds wonderful, but for IOT you need a processor with a very large production run to get the cost down. Maybe someday SIMD single chip solutions will be there, but not in the near term.
3) Reversible computers are delightful ideas, but there are fundamental issues running them at room temperature.
Because of the kT/q slopes in diodes, you can't actually recover the energy. Hardware was made 20 years ago and shows that nicely. Now at mK temperatures, that is a different story.
--R
On Wed, Jul 6, 2016 at 11:48 AM Henry Baker <hbaker1@pipeline.com> wrote:
Yes, but...
1. Memory is cheap. Even uP will have decent sized memories. Every delta in line widths produces at least O(delta^2) in memory size, if not O(delta^3).
2. SIMD-type processors are cheap and quite energy-efficient. There's really no reason for a traditional 16-bit processor anymore.
3. Most crypto codes are 1-1 functions, and therefore *reversible*. In theory, reversible functions can be computed with asymptotically zero energy dissipation.
At 07:31 AM 7/6/2016, Richard Howard wrote:
BTW, there is an interesting mathy problem here--how do you do security when you have only a few microjoules and a few kB of memory in a 16 bit processor?
RSA is not even remotely feasible.
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
Er --- like ... brains? WFL On 7/7/16, Richard Howard <rich@richardehoward.com> wrote:
Superconducting have some serious limitations, of which temperature is only one.
They have significant ac loss from flux motion and the long penetration depths associated with high temperature materials means that they are not good at confining magnetic fields--as is necessary for small inductors and transmission lines. Optical interconnects would be good alternatives, if low energy photons weren't so big and coupling strengths so small.
Quantum computation sounds wonderful (in the original sense of "wonder"), but unlikely to be useful in leaf nodes of the IOT in my lifetime. Perhaps we can use molecular-based QC using biological quantum effects, like seen in chlorophyll and grow the devices.
--R
On Wed, Jul 6, 2016 at 5:42 PM Tom Knight <tk@mit.edu> wrote:
If you have superconducting interconnect, then good inductors are possible. If you can make good inductors, you can make high speed reversible circuits easily (you are not limited to adiabatic processing). There still is a huge advantage to being able to have small amounts of irreversible logic (that is, to throw away some information). I think of it in the same vein as functional programming. You want to build programs in a functional style except for a very few places, where it really really pays off to have side effects.
I have been pushing DARPA to get serious about superconducting on-chip interconnect for about 20 years. In addition to good inductors (and hence high speed reversible circuits) it also speeds up on-chip communication from about 0.03c to about 0.6c. My more cynical view is that there is already such a program, but that it is classified. Along with nano-vacuum tubes.
On Jul 6, 2016, at 2:43 PM, Henry Baker <hbaker1@pipeline.com> wrote:
I haven't followed the "reversible" research for several years now, but I think that this research somehow they got off on the wrong track to begin with, and never quite got back.
In the '90's, the goal was to *lower* the energy dissipation, but not eliminate it (for reversible circuits).
Thus, they did really slow logic, and some amount of charge recovery.
But -- as you point out -- if you're using a highly non-linear device like a diode -- there are severe limits on what can be achieved.
There needs to be some amount of non-linearity/error-correction in order to act as a refrigerator to reject the "heat" (in the form of noise) that intrudes into the computation. So we run into typical Maxwell's Demon issues.
So we need to move to computational elements which are more isolated from the environment to begin with -- i.e., these computations are more "insulated" from the environment.
It so happens that the requirements for reversible computation overlap those for quantum computations almost completely. After all, quantum computations are reversible, so they (using non-quantum modes) can emulate classical reversible computations.
The real question is: if you have to do many of the same things for reversible computations as for quantum computations, then why not go directly to quantum computations -- which are strictly more powerful?
That's probably why most non-quantum reversible computation research stopped a number of years ago.
Now one aspect of reversible computation research hasn't (to my knowledge) been worked on very much: the emulation of complex reversible circuits using compositions of simpler reversible elements.
Sooner or later, such compositions will be required for quantum computations, as well, but I guess much of this research has been tabled until actual quantum computers can be built. This is a pity, as it would be interesting to better understand the *synthesis* of complex reversible circuits from elementary reversible gates. Such an understanding may provide more insight into why certain reversible circuits are *hard to invert* -- e.g., encryption circuits.
At 10:24 AM 7/6/2016, Richard Howard wrote:
1. Cheap means something different if you are talking about ubiquitous IOT nodes.
Large amount of embedded memory are very expensive and external RAM chips really break the bank for cost/size etc.
Trend is to single chip solutions with MCU/radio/memory. Not compatible with large quantities of DRAM.
As to delta in linewidth--the news is that the end is near because we are running out of atoms. Check the ITRS roadmap and Intel.
I have a nice photo of a CMOS device with 142 atoms from source to drain and 6 atoms of oxide (you can see the dots).
That was 20 years ago in research--we are there on-schedule in production.
2. Sounds wonderful, but for IOT you need a processor with a very large production run to get the cost down. Maybe someday SIMD single chip solutions will be there, but not in the near term.
3) Reversible computers are delightful ideas, but there are fundamental issues running them at room temperature.
Because of the kT/q slopes in diodes, you can't actually recover the energy. Hardware was made 20 years ago and shows that nicely. Now at mK temperatures, that is a different story.
--R
On Wed, Jul 6, 2016 at 11:48 AM Henry Baker <hbaker1@pipeline.com> wrote:
Yes, but...
1. Memory is cheap. Even uP will have decent sized memories. Every delta in line widths produces at least O(delta^2) in memory size, if not O(delta^3).
2. SIMD-type processors are cheap and quite energy-efficient. There's really no reason for a traditional 16-bit processor anymore.
3. Most crypto codes are 1-1 functions, and therefore *reversible*. In theory, reversible functions can be computed with asymptotically zero energy dissipation.
At 07:31 AM 7/6/2016, Richard Howard wrote:
BTW, there is an interesting mathy problem here--how do you do security when you have only a few microjoules and a few kB of memory in a 16 bit processor?
RSA is not even remotely feasible.
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
Indeed. See the stories of Cordwainer Smith (d1966). What goes around comes around. --R On Thu, Jul 7, 2016 at 11:26 AM Fred Lunnon <fred.lunnon@gmail.com> wrote:
Er --- like ... brains? WFL
On 7/7/16, Richard Howard <rich@richardehoward.com> wrote:
Superconducting have some serious limitations, of which temperature is only one.
They have significant ac loss from flux motion and the long penetration depths associated with high temperature materials means that they are not good at confining magnetic fields--as is necessary for small inductors and transmission lines. Optical interconnects would be good alternatives, if low energy photons weren't so big and coupling strengths so small.
Quantum computation sounds wonderful (in the original sense of "wonder"), but unlikely to be useful in leaf nodes of the IOT in my lifetime. Perhaps we can use molecular-based QC using biological quantum effects, like seen in chlorophyll and grow the devices.
--R
On Wed, Jul 6, 2016 at 5:42 PM Tom Knight <tk@mit.edu> wrote:
If you have superconducting interconnect, then good inductors are possible. If you can make good inductors, you can make high speed reversible circuits easily (you are not limited to adiabatic processing). There still is a huge advantage to being able to have small amounts of irreversible logic (that is, to throw away some information). I think of it in the same vein as functional programming. You want to build programs in a functional style except for a very few places, where it really really pays off to have side effects.
I have been pushing DARPA to get serious about superconducting on-chip interconnect for about 20 years. In addition to good inductors (and hence high speed reversible circuits) it also speeds up on-chip communication from about 0.03c to about 0.6c. My more cynical view is that there is already such a program, but that it is classified. Along with nano-vacuum tubes.
On Jul 6, 2016, at 2:43 PM, Henry Baker <hbaker1@pipeline.com> wrote:
I haven't followed the "reversible" research for several years now, but I think that this research somehow they got off on the wrong track to begin with, and never quite got back.
In the '90's, the goal was to *lower* the energy dissipation, but not eliminate it (for reversible circuits).
Thus, they did really slow logic, and some amount of charge recovery.
But -- as you point out -- if you're using a highly non-linear device like a diode -- there are severe limits on what can be achieved.
There needs to be some amount of non-linearity/error-correction in order to act as a refrigerator to reject the "heat" (in the form of noise) that intrudes into the computation. So we run into typical Maxwell's Demon issues.
So we need to move to computational elements which are more isolated from the environment to begin with -- i.e., these computations are more "insulated" from the environment.
It so happens that the requirements for reversible computation overlap those for quantum computations almost completely. After all, quantum computations are reversible, so they (using non-quantum modes) can emulate classical reversible computations.
The real question is: if you have to do many of the same things for reversible computations as for quantum computations, then why not go directly to quantum computations -- which are strictly more powerful?
That's probably why most non-quantum reversible computation research stopped a number of years ago.
Now one aspect of reversible computation research hasn't (to my knowledge) been worked on very much: the emulation of complex reversible circuits using compositions of simpler reversible elements.
Sooner or later, such compositions will be required for quantum computations, as well, but I guess much of this research has been tabled until actual quantum computers can be built. This is a pity, as it would be interesting to better understand the *synthesis* of complex reversible circuits from elementary reversible gates. Such an understanding may provide more insight into why certain reversible circuits are *hard to invert* -- e.g., encryption circuits.
At 10:24 AM 7/6/2016, Richard Howard wrote:
1. Cheap means something different if you are talking about ubiquitous IOT nodes.
Large amount of embedded memory are very expensive and external RAM chips really break the bank for cost/size etc.
Trend is to single chip solutions with MCU/radio/memory. Not compatible with large quantities of DRAM.
As to delta in linewidth--the news is that the end is near because we are running out of atoms. Check the ITRS roadmap and Intel.
I have a nice photo of a CMOS device with 142 atoms from source to drain and 6 atoms of oxide (you can see the dots).
That was 20 years ago in research--we are there on-schedule in production.
2. Sounds wonderful, but for IOT you need a processor with a very large production run to get the cost down. Maybe someday SIMD single chip solutions will be there, but not in the near term.
3) Reversible computers are delightful ideas, but there are fundamental issues running them at room temperature.
Because of the kT/q slopes in diodes, you can't actually recover the energy. Hardware was made 20 years ago and shows that nicely. Now at mK temperatures, that is a different story.
--R
On Wed, Jul 6, 2016 at 11:48 AM Henry Baker <hbaker1@pipeline.com> wrote:
Yes, but...
1. Memory is cheap. Even uP will have decent sized memories. Every delta in line widths produces at least O(delta^2) in memory size, if not O(delta^3).
2. SIMD-type processors are cheap and quite energy-efficient. There's really no reason for a traditional 16-bit processor anymore.
3. Most crypto codes are 1-1 functions, and therefore *reversible*. In theory, reversible functions can be computed with asymptotically zero energy dissipation.
At 07:31 AM 7/6/2016, Richard Howard wrote: > BTW, there is an interesting mathy problem here--how do you do security > when you have only a few microjoules and a few kB of memory in a 16 bit > processor? > > RSA is not even remotely feasible.
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun
participants (4)
-
Fred Lunnon -
Henry Baker -
Richard Howard -
Tom Knight