It's so ambitious and so questionable at the same time. I'd classify it as hard science fiction, like a really good Orion's Arm entry, if the authors weren't from NIST.
Final two discussion paragraphs from paper V:
While it may be difficult to build systems larger than 10 billion neurons in the near term, such a system is not physically limited. Like the brain, such limits will be incurred due to the velocity of signal propagation. From Fig. 6(c) we know that networks as large as data centers can sustain coherent oscillations at 1 MHz. Such a facility would house 10^8 300 mm wafers if they were stacked 100 deep. This would result in 100 trillion neurons per data center across modules interconnected with another power law distribution.
Networks need not oscillate at 1 MHz, and if they supported system-wide activity at 1 kHz—faster than any oscillation of the human brain—the neuronal pool could occupy a significant fraction of the earth’s surface and employ quintillions of neurons. We do not wish to cover earth in such devices, but asteroids provide ample, uncontroversial real estate. The materials for this hardware are abundant on M-type and S-type asteroids [76–80]. It appears possible for an asteroid belt to form the nodes and light to form the edges of a solar-system scale intelligent network. Asteroids can be separated by billions of meters, so light-speed communication delays may be several seconds or longer. For cognitive systems oscillating up to 20 MHz, such delays would cause individual modules to operate as separate cognitive systems, much like a society of humans.
Apart from the breathtaking scale of speculation -- which one could admittedly also find to good effect in older papers about e.g. nuclear power -- there is a more concrete question. What's the all-in energy cost-per-operation vs. conventional hardware, CMOS devices operating above room temperature? Operating at liquid helium temperatures dramatically shrinks the on-chip power demand, and then the cryocooler dramatically re-inflates it. Lab scale production of liquid helium takes ~570 watts of wall-plug power to produce 1 watt of cooling near 4.2 K [1]. At the wall plug, cryogenic cooling systems included, how does this design compare to existing hardware on power and speed for neural network training or inference? AFAICT, the authors do not attempt to estimate such a figure of merit.
I answered more of my questions about energy costs: in part II Appendix C the authors explore the energy of a synaptic firing event. The on-chip energy expended: 41 attojoules. (Though in part V the authors conclude that their earlier Part II analysis was probably flawed, and the circuits probably need 4 times as much current.) That 41 attojoules on-chip translates to about 24 femtojoules of whole-system energy consumption. That in turn compares favorably to the ~10 picojoules (10000 femtojoule) of energy required to perform a single precision floating point arithmetic operation on CMOS hardware. But this "neuromorphic" design is not easy to compare to contemporary hardware on complete tasks, because it's more analog than digital. And it's not clear that it can easily interoperate with existing artificial neural networks built on digital logic, either for training or inference. A low level "direct" comparison may be effectively impossible. Instead we'd have to ask questions like "how many joules per face recognized?" for complete facial recognition systems, admitting that GPU-based ANNs and superconducting optoelectronic ANNs would have very different internal structures.
Finally, I'll note that these designs rely on large numbers of Josephson junctions. From my quick Wiki-skim it appears that nobody has ever built large scale integrated circuits from Josephson junctions. It doesn't look like there have even been serious Western attempts at it after the 1980s. That's not to say that large scale fabrication of circuits incorporating Josephson junctions is folly, but it looks like it requires a lot of basic R&D effort before somebody can build one of these superconducting optoelectronic chips, much less a whole wafer full of interconnected copies.
I don't think a star-system sized computational network is supposed to be used for the mere face recognition that my phone can already do. So the metric may be comparing apples to oranges.
Pure speculation, but one possible benefit of dramatically reducing the on-chip power requirements would be allowing larger areas of the chip to be active for longer periods of time. Most of the area of a cpu is inactive during execution, or they would melt...
Being able to have stacked silicon neural network operate at peak speed everywhere could be worth the tremendous cooling costs.
Not even remotely. For one thing, it's not even that cold out there; average temperature out at the main belt is only around -75 C, based on radiative equilibrium with the incoming sunlight. (Obviously, the near-earth asteroids are warmer.) And then you've got the problem that you're surrounded by vacuum, so the only way you have to dump heat for further cooling is through thermal radiation, which is terribly inefficient. Not to mention, thirdly, the fact that there's no natural helium available on asteroids, so you'd have to ship it up constantly to replace your coolant as it boiled off.
Meanwhile, here on Earth, we have billions of cubic feet of helium gas underground with infrastructure for liquefying it, we have access to conductive and convective cooling for our heat pumps with the whole thermal mass of the Earth to sink heat into, and we have all the industrial infrastructure and construction and maintenance infrastructure right here.
Oh, and if your datacenter is in Palo Alto instead of on Pallas, you have the added bonus that you don't have to worry about networking with a 15-minute ping.
tl;dr space isn't cold, vacuum isn't a great cooling medium, and shipping is expensive.
You'd have your radiators in the shade, where it's at least a hundred degrees cooler and you have the thermal mass of the asteroid where you can dump heat temporarily. Sourcing Helium might be difficult, but maybe you can get it from one of the gas giants?
That'd be even less efficient; the gas giants have much deeper gravity wells than Earth, and it's correspondingly much harder to bring material up from them. Jupiter's exhaust velocity is 60 km/s, Saturn's is 35 - compared to a measly 11 km/s for Earth. And as the rocket equation tells us, the fuel required goes up exponentially the greater the change in velocity you need.
This is easy to google (I should have done so before) but still hard to believe. If that's peak speed for only a short time because of fuel limits, it's less hard to believe.
...what do you mean by "peak speed for only a short time because of fuel limits?" It's space. You don't need fuel to maintain speed. What's going to slow you down?
Oh: To clarify, "Escape velocity" is the speed you need to escape Earth's gravity well entirely, so that you are no longer orbiting the Earth at all; it's about the amount of velocity you need to go on interplanetary missions. If you just want to go to, say, the Moon, then you don't need to go that fast. If you just want to go to low Earth orbit (LEO), then orbital velocity's only around 7 km/s.
But that isn't "peak speed" or anything, that's just ... how fast you have to go to be in low Earth orbit. The International Space Station is traveling at 7.67 km/s right now.
Superconducting Optoelectronic Neurons II: Receiver Circuits: https://arxiv.org/abs/1805.02599
Superconducting Optoelectronic Neurons III: Synaptic Plasticity: https://arxiv.org/abs/1805.01937
Superconducting Optoelectronic Neurons IV: Transmitter Circuits: https://arxiv.org/abs/1805.01941
Superconducting Optoelectronic Neurons V: Networks and Scaling: https://arxiv.org/abs/1805.01942
It's so ambitious and so questionable at the same time. I'd classify it as hard science fiction, like a really good Orion's Arm entry, if the authors weren't from NIST.
Final two discussion paragraphs from paper V:
While it may be difficult to build systems larger than 10 billion neurons in the near term, such a system is not physically limited. Like the brain, such limits will be incurred due to the velocity of signal propagation. From Fig. 6(c) we know that networks as large as data centers can sustain coherent oscillations at 1 MHz. Such a facility would house 10^8 300 mm wafers if they were stacked 100 deep. This would result in 100 trillion neurons per data center across modules interconnected with another power law distribution.
Networks need not oscillate at 1 MHz, and if they supported system-wide activity at 1 kHz—faster than any oscillation of the human brain—the neuronal pool could occupy a significant fraction of the earth’s surface and employ quintillions of neurons. We do not wish to cover earth in such devices, but asteroids provide ample, uncontroversial real estate. The materials for this hardware are abundant on M-type and S-type asteroids [76–80]. It appears possible for an asteroid belt to form the nodes and light to form the edges of a solar-system scale intelligent network. Asteroids can be separated by billions of meters, so light-speed communication delays may be several seconds or longer. For cognitive systems oscillating up to 20 MHz, such delays would cause individual modules to operate as separate cognitive systems, much like a society of humans.
Apart from the breathtaking scale of speculation -- which one could admittedly also find to good effect in older papers about e.g. nuclear power -- there is a more concrete question. What's the all-in energy cost-per-operation vs. conventional hardware, CMOS devices operating above room temperature? Operating at liquid helium temperatures dramatically shrinks the on-chip power demand, and then the cryocooler dramatically re-inflates it. Lab scale production of liquid helium takes ~570 watts of wall-plug power to produce 1 watt of cooling near 4.2 K [1]. At the wall plug, cryogenic cooling systems included, how does this design compare to existing hardware on power and speed for neural network training or inference? AFAICT, the authors do not attempt to estimate such a figure of merit.
[1] Basics of low-temperature refrigeration: https://cds.cern.ch/record/1974048/files/arXiv:1501.07392.pd...