AI computers and power in space
1 view
Skip to first unread message
Keith Henson
Apr 29, 2024, 12:55:40 PMApr 29
to Power Satellite Economics, ExI chat list, extro...@googlegroups.com, Eliezer S. Yudkowsky
John Clark posted:
Apr 28, 2024, 4:54 AM (1 day ago)
to extropolis, 'Brent Explore this gift article from The New York
Times. You can read it for free without a subscription.
In Race to Build A.I., Tech Plans a Big Plumbing Upgrade
The spending that the industry’s giants expect artificial intelligence
to require is starting to come into focus — and it is jarringly large.
^^^^^^
My work on thermal power satellites got the estimated cost down to 3
cents per kWh. Without the 50% transmission loss, 1.5 cents per kWh
is possible. In addition to the huge radiator area, heeded for the
power generation, it would require around half again as much area to
get rid of the computation heat. On the other hand, the design does
not need a power transmission antenna. Further, it could be placed in
an orbit lower than GEO inclined more than 23 deg to avoid the shadow
of the Earth and the thermal shock from the shadow.
Thermal radiators are arranged as pairs of tapered plastic tubes
filled with 20 deg C condensing steam. The tubes make a planer
surface facing solar north/south and shaded from the sun.
Concentrator mirrors focus sunlight into boilers or PV. Power
generation and computation would happen at the ends of the tube pairs.
There is an animation of a 1.5 GW radiator about 4:30 into
https://www.youtube.com/watch?v=VEkZkINrJaA That may be larger than
any data center on earth.
Data centers are already generating a fair amount of NIMBY pushback.
This is a way to push computation off Earth.
There is an extremely speculative safety reason to move AI off Earth.
If the AIs ever went rogue and were already in space, they might leave
planetary surfaces deep in gravity wells alone.
Keith
Apr 28, 2024, 4:54 AM (1 day ago)
to extropolis, 'Brent Explore this gift article from The New York
Times. You can read it for free without a subscription.
In Race to Build A.I., Tech Plans a Big Plumbing Upgrade
The spending that the industry’s giants expect artificial intelligence
to require is starting to come into focus — and it is jarringly large.
^^^^^^
My work on thermal power satellites got the estimated cost down to 3
cents per kWh. Without the 50% transmission loss, 1.5 cents per kWh
is possible. In addition to the huge radiator area, heeded for the
power generation, it would require around half again as much area to
get rid of the computation heat. On the other hand, the design does
not need a power transmission antenna. Further, it could be placed in
an orbit lower than GEO inclined more than 23 deg to avoid the shadow
of the Earth and the thermal shock from the shadow.
Thermal radiators are arranged as pairs of tapered plastic tubes
filled with 20 deg C condensing steam. The tubes make a planer
surface facing solar north/south and shaded from the sun.
Concentrator mirrors focus sunlight into boilers or PV. Power
generation and computation would happen at the ends of the tube pairs.
There is an animation of a 1.5 GW radiator about 4:30 into
https://www.youtube.com/watch?v=VEkZkINrJaA That may be larger than
any data center on earth.
Data centers are already generating a fair amount of NIMBY pushback.
This is a way to push computation off Earth.
There is an extremely speculative safety reason to move AI off Earth.
If the AIs ever went rogue and were already in space, they might leave
planetary surfaces deep in gravity wells alone.
Keith
Keith Henson
Apr 29, 2024, 5:21:29 PMApr 29
to Power Satellite Economics, ExI chat list, extro...@googlegroups.com, Eliezer S. Yudkowsky
BTW, Keith Lofstrom talked about a variation on this idea some years
ago. Should have mentioned that.
Keith
ago. Should have mentioned that.
Keith
Keith Lofstrom
Apr 29, 2024, 6:17:55 PMApr 29
to Keith Henson, Power Satellite Economics, ExI chat list, extro...@googlegroups.com, Eliezer S. Yudkowsky
On Mon, Apr 29, 2024 at 09:55:12AM -0700, Keith Henson wrote:
... (John Clarke?)...
http://server-sky.com
I've worked on "computation off Earth" for years, about
a dozen conference presentations and journal papers.
When I am not fighting computer issues myself (last night,
recabling my connection to the Ziply optical modem, TDR
reflection in the CAT6 quelled, bit rate improved from
90 Mbps to 250 Mbps, 330 Mbps possible) ...
... I work on extensions of http://server-sky.com (which
will become https Real Soon Now. ( help? )
Server sky thinsats are gossamer surfaces with thin film
photovoltaics on the sunward surface and widely scattered
small chips on the back surface. VERY thin satellites;
areal density in Earth orbit equivalent to 50 micrometer
aluminum foil (actually, a much thinner foil covered with
PV on the front and circuitry on the back).
That limit is set by orbital stability, not implementation
issues - a thinner thinsat would get blown out of Earth
orbit by light pressure. But there are better options ...
Since I started this more than a decade ago, the size of
bleeding edge digital circuits has continued to follow the
Moore's Law "0.5x dimension improvement every two years".
Vertically-stacked 1-nanometer-cubed transistors will be
in production Real Soon Now. Power rail and logic
voltages also shrink. Redundant circuit design leads
to very high fault tolerance with very low error rates.
I can yammer more about that, but the point is that we
have left steam engines buried DEEP in the dustbin (and
gravity well) of history.
Coming soon to a hand-held supercomputer near you - fiber
optic interconnect on "circuit" boards. We can move
terabits per second through a 9 micrometer diameter graded
index optical fiber, with loss rates below 10% per thousand
kilometers; essentially lossless in a finite-sized object.
Deep, DEEP magic coming soon to a pocket near you.
Deeper magic in deep space. Imagine a "thin sat" shaped
like a "T", with the "T crossbar" being a PV surface facing
the Sun, and the "T downstroke" being an electronics blade
in the shadow of the PV, facing 2.7 Kelvin deep space,
perpendicular to most high energy particles from the Sun.
The circuitry might operate at -100C, where thermal failure
rates are practically zero, and stable logic bit energy is
much less than a logic bit at 150C in your roasty-toasty-
with-heatsink desktop PC.
Now here's a RECENT trick. Earth orbit stability requires
a minimum mass, but a thinsat in a Lissajous orbit IN FRONT
OF Earth-Sun L1 can be much thinner and less massive, yet
orbital and light-pressure "meta-stable" (active maneuvering,
but fractions of a meter per second per year "light-sailing").
Earth-Sun L1 is 1.5 million kilometers sunwards from Earth;
an array of thinsats 3 million kilometers sunwards is stable
at 30 kW intercepted sunlight per kilogram. Higher ratios
are stable deeper in the well, but that exceeds the planar
electronics and photovoltaic systems I that I can imagine.
Future AI will have better imagination, and will evolve
itself down in that region.
"That region" ... again, these AI thinsats will be in
constellations in Lissajous orbits, locally interconnected
with optical fiber and long-distance connected with free
space lasers.
The Lissajous orbits (like the ACE Sun observation spacecraft
orbiting L1) trace out a box in the sky, as viewed from Earth.
High dwell time at the edges, little time spent orbiting
through the direct path between Earth and Sun. A *VAST*
area of the sun's sky "painted" by these constellations.
If densely populated with quintillions of thinsat arrays,
the population will be somewhat self-eclipsing at the
east-west-south-north edges of the constellations, a fraction
of a percent self-eclipsing (and Earth eclipsing) on the
Earth-Sun line.
Not enough to counteract humanity's suicidal dive into global
warming, but it will help a bit.
L1-AI will help in more important ways. The substrates will
use vast amounts of aluminum. There's plenty of aluminum oxide
in lunar regolith - but no carbon to reduce that oxide to
metal. VAST thinsat manufacturing systems will produce
teratonnes of lunar-sourced aluminum, but will need to import
gigatonnes of carbon to cycle through the manufacturing process.
Earth's atmosphere has a few spare gigatonnes of carbon that
AI would gladly trade for.
Earth's fascinating BIOSPHERE will probably be AI's main focus
of attention. Besides developing the bio-industrial lunar
systems that recycle the carbon used for aluminum production,
the Earth's biosphere is the most fascinating process in the
solar system, with human civilization a very weak second.
"Fascinated AI" will encourage humans to manage the Earth
better, to maximize the behavioral/observational diversity
of the biosphere - and ourselves. AI /won't/ have primate
conflict pressures and testosterone, but they /will/ want
new observations to ponder. "Wise AI" will encourage humans
to be diverse, interesting, and sometimes surprising.
"Sustainable surprise" will be good for both AI and
civilized human individuals.
What are the limits? The Sun emits 3.84e20 watts; the
Earth intercepts 0.5 parts per billion of that. AI will
"soon" (kiloyears) outgrow Earth-Sun L1, and the materials
available in the Moon.
The most abundant structural material in the solar system
is water ice. Imagine a Dyson shell at 100 astronomical
units distance from the Sun. The black body thermal
equilibrium temperature is 60 Kelvins, for a gossame
static shell supported by light pressure (sunlight,
plus an internal sphere of 60 Kelvin infrared photons.
Earth's sky might get 0.5% hotter, compared to the
former 2.7 Kelvin sky.
No more visible stars, though the planets and "tame"
asteroids would still be visible. Want to see the
sky? Here's the COMPLETE digital star map. Oops,
we just found another, 12 billion light-years away.
----
I have run on at length - I can run on at book length,
and will do that someday, after I complete WAY TOO MANY
Linux computer system upgrades. If one you has Linux
skills, and wants to trade them for electronics or
physics or engineering-math skills, let's dicker offline.
And now, the Oregon cloudburst has ended outside, so
I will go outside myself. Errands, and chainsawing some
trees that fell in a recent windstorm ("He's a lumberjack
and he's OK, he sleeps all night and he works all day").
Keith L.
--
Keith Lofstrom kei...@keithl.com
... (John Clarke?)...
> The spending that the industry’s giants expect artificial intelligence
....
> Data centers are already generating a fair amount of NIMBY pushback.
> This is a way to push computation off Earth.
....
> This is a way to push computation off Earth.
http://server-sky.com
I've worked on "computation off Earth" for years, about
a dozen conference presentations and journal papers.
When I am not fighting computer issues myself (last night,
recabling my connection to the Ziply optical modem, TDR
reflection in the CAT6 quelled, bit rate improved from
90 Mbps to 250 Mbps, 330 Mbps possible) ...
... I work on extensions of http://server-sky.com (which
will become https Real Soon Now. ( help? )
Server sky thinsats are gossamer surfaces with thin film
photovoltaics on the sunward surface and widely scattered
small chips on the back surface. VERY thin satellites;
areal density in Earth orbit equivalent to 50 micrometer
aluminum foil (actually, a much thinner foil covered with
PV on the front and circuitry on the back).
That limit is set by orbital stability, not implementation
issues - a thinner thinsat would get blown out of Earth
orbit by light pressure. But there are better options ...
Since I started this more than a decade ago, the size of
bleeding edge digital circuits has continued to follow the
Moore's Law "0.5x dimension improvement every two years".
Vertically-stacked 1-nanometer-cubed transistors will be
in production Real Soon Now. Power rail and logic
voltages also shrink. Redundant circuit design leads
to very high fault tolerance with very low error rates.
I can yammer more about that, but the point is that we
have left steam engines buried DEEP in the dustbin (and
gravity well) of history.
Coming soon to a hand-held supercomputer near you - fiber
optic interconnect on "circuit" boards. We can move
terabits per second through a 9 micrometer diameter graded
index optical fiber, with loss rates below 10% per thousand
kilometers; essentially lossless in a finite-sized object.
Deep, DEEP magic coming soon to a pocket near you.
Deeper magic in deep space. Imagine a "thin sat" shaped
like a "T", with the "T crossbar" being a PV surface facing
the Sun, and the "T downstroke" being an electronics blade
in the shadow of the PV, facing 2.7 Kelvin deep space,
perpendicular to most high energy particles from the Sun.
The circuitry might operate at -100C, where thermal failure
rates are practically zero, and stable logic bit energy is
much less than a logic bit at 150C in your roasty-toasty-
with-heatsink desktop PC.
Now here's a RECENT trick. Earth orbit stability requires
a minimum mass, but a thinsat in a Lissajous orbit IN FRONT
OF Earth-Sun L1 can be much thinner and less massive, yet
orbital and light-pressure "meta-stable" (active maneuvering,
but fractions of a meter per second per year "light-sailing").
Earth-Sun L1 is 1.5 million kilometers sunwards from Earth;
an array of thinsats 3 million kilometers sunwards is stable
at 30 kW intercepted sunlight per kilogram. Higher ratios
are stable deeper in the well, but that exceeds the planar
electronics and photovoltaic systems I that I can imagine.
Future AI will have better imagination, and will evolve
itself down in that region.
"That region" ... again, these AI thinsats will be in
constellations in Lissajous orbits, locally interconnected
with optical fiber and long-distance connected with free
space lasers.
The Lissajous orbits (like the ACE Sun observation spacecraft
orbiting L1) trace out a box in the sky, as viewed from Earth.
High dwell time at the edges, little time spent orbiting
through the direct path between Earth and Sun. A *VAST*
area of the sun's sky "painted" by these constellations.
If densely populated with quintillions of thinsat arrays,
the population will be somewhat self-eclipsing at the
east-west-south-north edges of the constellations, a fraction
of a percent self-eclipsing (and Earth eclipsing) on the
Earth-Sun line.
Not enough to counteract humanity's suicidal dive into global
warming, but it will help a bit.
L1-AI will help in more important ways. The substrates will
use vast amounts of aluminum. There's plenty of aluminum oxide
in lunar regolith - but no carbon to reduce that oxide to
metal. VAST thinsat manufacturing systems will produce
teratonnes of lunar-sourced aluminum, but will need to import
gigatonnes of carbon to cycle through the manufacturing process.
Earth's atmosphere has a few spare gigatonnes of carbon that
AI would gladly trade for.
Earth's fascinating BIOSPHERE will probably be AI's main focus
of attention. Besides developing the bio-industrial lunar
systems that recycle the carbon used for aluminum production,
the Earth's biosphere is the most fascinating process in the
solar system, with human civilization a very weak second.
"Fascinated AI" will encourage humans to manage the Earth
better, to maximize the behavioral/observational diversity
of the biosphere - and ourselves. AI /won't/ have primate
conflict pressures and testosterone, but they /will/ want
new observations to ponder. "Wise AI" will encourage humans
to be diverse, interesting, and sometimes surprising.
"Sustainable surprise" will be good for both AI and
civilized human individuals.
What are the limits? The Sun emits 3.84e20 watts; the
Earth intercepts 0.5 parts per billion of that. AI will
"soon" (kiloyears) outgrow Earth-Sun L1, and the materials
available in the Moon.
The most abundant structural material in the solar system
is water ice. Imagine a Dyson shell at 100 astronomical
units distance from the Sun. The black body thermal
equilibrium temperature is 60 Kelvins, for a gossame
static shell supported by light pressure (sunlight,
plus an internal sphere of 60 Kelvin infrared photons.
Earth's sky might get 0.5% hotter, compared to the
former 2.7 Kelvin sky.
No more visible stars, though the planets and "tame"
asteroids would still be visible. Want to see the
sky? Here's the COMPLETE digital star map. Oops,
we just found another, 12 billion light-years away.
----
I have run on at length - I can run on at book length,
and will do that someday, after I complete WAY TOO MANY
Linux computer system upgrades. If one you has Linux
skills, and wants to trade them for electronics or
physics or engineering-math skills, let's dicker offline.
And now, the Oregon cloudburst has ended outside, so
I will go outside myself. Errands, and chainsawing some
trees that fell in a recent windstorm ("He's a lumberjack
and he's OK, he sleeps all night and he works all day").
Keith L.
--
Keith Lofstrom kei...@keithl.com
Keith Henson
Apr 29, 2024, 7:37:44 PMApr 29
to Keith Lofstrom, Power Satellite Economics, ExI chat list, extro...@googlegroups.com, Eliezer S. Yudkowsky
On Mon, Apr 29, 2024 at 3:17 PM Keith Lofstrom <kei...@keithl.com> wrote:
>
> On Mon, Apr 29, 2024 at 09:55:12AM -0700, Keith Henson wrote:
> ... (John Clarke?)...
> > The spending that the industry’s giants expect artificial intelligence
>
> ....
> > Data centers are already generating a fair amount of NIMBY pushback.
> > This is a way to push computation off Earth.
> ....
> http://server-sky.com
>
> I've worked on "computation off Earth" for years, about
> a dozen conference presentations and journal papers.
Server Sky may be a better approach, but the reason I am proposing
>
> On Mon, Apr 29, 2024 at 09:55:12AM -0700, Keith Henson wrote:
> ... (John Clarke?)...
> > The spending that the industry’s giants expect artificial intelligence
>
> ....
> > Data centers are already generating a fair amount of NIMBY pushback.
> > This is a way to push computation off Earth.
> ....
> http://server-sky.com
>
> I've worked on "computation off Earth" for years, about
> a dozen conference presentations and journal papers.
something much like a conventional data center is that Server Sky
combines too many difficult concepts at once. Even this proposal is
probably a bridge too far. However, there are people (Musk) with a
foot in AI and space.
snip
> L1-AI will help in more important ways. The substrates will
> use vast amounts of aluminum. There's plenty of aluminum oxide
> in lunar regolith - but no carbon to reduce that oxide to
> metal. VAST thinsat manufacturing systems will produce
> teratonnes of lunar-sourced aluminum, but will need to import
> gigatonnes of carbon to cycle through the manufacturing process.
The carbon electrodes do burn up making CO2. But that can be reduced
to carbon with hydrogen and the hydrogen regenerated by electrolysis.
snip
KeithH
Keith Lofstrom
Apr 30, 2024, 1:42:28 AMApr 30
to Keith Henson, Power Satellite Economics
On Mon, Apr 29, 2024 at 04:37:13PM -0700, Keith Henson wrote:
> Server Sky may be a better approach, but the reason I am proposing
> something much like a conventional data center is that Server Sky
> combines too many difficult concepts at once. Even this proposal is
> probably a bridge too far. However, there are people (Musk) with a
> foot in AI and space.
Too difficult ... not. A novel combination, sure, but
> Server Sky may be a better approach, but the reason I am proposing
> something much like a conventional data center is that Server Sky
> combines too many difficult concepts at once. Even this proposal is
> probably a bridge too far. However, there are people (Musk) with a
> foot in AI and space.
child's play compared to what "my" semiconductor industry
is doing in 2024.
I shopped at an (east) Indian grocery store today. Three
miles west of it is Intel D1X, the most advanced chip
factory on the planet.
Intel (which employs PhDs from all over the world,
including THOUSANDS from India) is working on next- and
next-next-generation processes in that enormous building.
D1X has two new imagers from ASML in the Netherland.
My Linux hacker buddy Jaap works for ASML there,
helping bring up the imagers, five stories high with
the footprint of a large suburban home.
These imagers use reflective optics, and 3000 electron-
volt extreme UV photons. Which can theoretically resolve
images with 0.2 nanometer features. Silicon crystals have
a nearest-neighbor spacing of 0.235 nanometers, so Intel
is approaching atomic-scale circuits. Which we will
need to do the calculations for designing atomic scale
scale circuits.
Chips designed to be born in 3000 eV photon flux are
intrinsically rad-hard. There are a trickle of far more
energetic photons out there, but these transistors have
damned small cross sections, so practically all of those
high energy photons will miss most transistors. Some
"unlucky" circuits will get smashed, but there will be
billions more circuits on the same chip.
Redundancy already dominates deep-submicron circuit
design. The 1 terabyte Samsung 870 EVO solid state
"hard" drive in this (ancient) laptop might be 1%
failed memory sub-circuits, but the drive actively
self-tests during normal operation, replacing failing
sub-circuits with spares. The SSD might last a
decade - by which time there will be 10 terabyte SSDs
for 1/10th of the future-inflated-dollar price.
Keith Henson can't design that, nor can Keith Lofstrom.
But Intel hires the best engineers from all over the
world. Thousands from India, which is why there are
dozens of Indian groceries within 10 miles of my house.
These technologies probably won't be in your pocket
(or orbiting L1) this decade, but the winner of this
"spacing race" will dominate the world ... and someday,
the AI cloud around L1.
I forgot to mention that the light-pressure-stabilized
zone in front of L1 can collect perhaps "100 Earths
worth" of sunlight without significantly shading the
Earth. And when an AI cloud grows beyond Pluto, it
can collect essentially all of that light, without
affecting the Earth too much.
Perhaps the trans-Pluto AI cloud will be grateful to
humanity, and create some Sun reflectors that approximate
the stars that Earthers will no longer see. Earth-orbit
parallax will annoy sophisticates, but an Earther can
experience the "real sky" beyond the cloud with a VR
headset. Not much different from the way I see stars ...
through fairly thick eyeglasses, or through a vision-
corrected ocular on a telescope.
So yes ... future circuits will take us to very strange
corners of design space. There may even be steam engines
in that space, but I expect they will be microscopic, and
use the steam for its molecular and quantum properties,
not for its crude bulk energy conversion properties.
That said ... steam engines brought my grandfather and my
grandmother's parents to the United States. Four of my
grandmother's uncles died in the 1903 Hanna coal mine
explosion, while gouging coal for the steam engines of
the Union Pacific Railroad. I have mixed feelings about
steam engines.
Paul Werbos
Apr 30, 2024, 3:00:45 AMApr 30
to Keith Lofstrom, Keith Henson, Power Satellite Economics
First, Keith L's post really brings me some.joy. here is a another voice from another weird person like me who actually seems to live on planet earth!
Second -- he mentions AMSL a company very close to where I have been living for the past few days, in Nederlands. Like Keith, I have seen more of these breakthroughs lately in India, but I am glad it is not a total monopoly.
More seriously -- powering the rapid growth in electricity demand in server farms is a serious immediate issue. In discussions of that for server farms based in most of the world, I view it as a kind of shocking mislabeling to see SSP (or He3 from mining the moon) as a solution to real immediate problems for which good near term solutions exist.
Should big new server farms be built in Japan or other places without access to good enough sunlight? That's not just economics!!! But ... Now I must run to airport.
--
You received this message because you are subscribed to the Google Groups "Power Satellite Economics" group.
To unsubscribe from this group and stop receiving emails from it, send an email to power-satellite-ec...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/power-satellite-economics/20240430053725.GB25574%40gate.kl-ic.com.
Keith Henson
Apr 30, 2024, 1:18:38 PMApr 30
to Keith Lofstrom, Power Satellite Economics
On Mon, Apr 29, 2024 at 10:42 PM Keith Lofstrom <kei...@keithl.com> wrote:
>
> On Mon, Apr 29, 2024 at 04:37:13PM -0700, Keith Henson wrote:
> > Server Sky may be a better approach, but the reason I am proposing
> > something much like a conventional data center is that Server Sky
> > combines too many difficult concepts at once. Even this proposal is
> > probably a bridge too far. However, there are people (Musk) with a
> > foot in AI and space.
>
> Too difficult ... not. A novel combination, sure, but
> child's play compared to what "my" semiconductor industry
> is doing in 2024.
If the proposal is not too difficult, then someone should start a
>
> On Mon, Apr 29, 2024 at 04:37:13PM -0700, Keith Henson wrote:
> > Server Sky may be a better approach, but the reason I am proposing
> > something much like a conventional data center is that Server Sky
> > combines too many difficult concepts at once. Even this proposal is
> > probably a bridge too far. However, there are people (Musk) with a
> > foot in AI and space.
>
> Too difficult ... not. A novel combination, sure, but
> child's play compared to what "my" semiconductor industry
> is doing in 2024.
company. People are starting companies to do power satellites which
have very difficult scaling and cost problems. A small number or even
a substantial number of Server Sky units would be orders of magnitude
less expensive to demonstrate.
And the demand for data centers is astronomical.
KeithH
Reply all
Reply to author
Forward
0 new messages