There’s a new computer chip called the Cerebras Wafer-Scale Engine which is massive.
Chips are usually the size of a postage stamp, and you get faster computers by linking a bunch of them together. A different approach: the Cerebras is the size of an iPad.
The Cerebras has 1.2 trillion transistors, 20 times bigger than the world’s next largest chip, and 75 times bigger than the new Apple M1 chip. Singularity Hub has all the details: The Trillion-Transistor Chip That Just Left a Supercomputer in the Dust.
Anyway, this line from the announcement caught my eye:
it can tell you what is going to happen in the future faster than the laws of physics produce that same result.
Faster than real-time simulation.
I mean, that’s what physics does anyway, right? We can figure out the position of the planets in the solar system for a thousand years in the future without having to run the calculation for a thousand years.
And yet… something provocative about this.
One of the challenges with nuclear fission, the holy grail of clear energy generating, is the plasma going out of control, and the magnetic fields in the torus can’t be adjusted fast enough to contain it. The physics is too hard; nobody’s “solved” the plasma problem yet. But if the entire thing can be simulated from the bottom up, faster than realtime, you don’t need a model. You just run the simulation.
What are the civilian applications?
I can imagine a wearable device that continuously snapshots the world around you, runs the simulation in fast forward, and pre-emptively warns you about a mugging, or a bike going off course. Call it augmented apprehension.
Or how about intelligent fire extinguishers that simulate the fire in faster-than-realtime and dynamically direct the spray to uncannily effective spots.
I think it’s that uncanniness that draws me the most. Fluid dynamics and chaotic systems generally are weird and interesting. I think of the weird interference patterns you get in a pool of water if you get ripples to meet up in specific ways, or the strange behaviours of inverted pendulums that stand upright if you vibrate them at the right frequency. (Human skeletons are basically realtime adjusted inverted pendulums.)
So, with powerful simulation, could you figure out how to hit a mass of water with puffs of air so that it rises up and moves around the room, washing the windows; or robots with reed-thin jointed limbs that should never be able to hold themselves up, but with motors at each joint running at just the right vibration to keep the thing moving?
The general algorithm would seem to be:
- sense and simulate the system
- solve for the most energy-efficient intervention that leads to a improbable yet desired outcome, faster than that outcome can occur in real time
- perform that action
- rinse and repeat
…which is how dolphins swim and bumblebees fly.
Just as machine learning is getting into everything, and changing all software to the point that we don’t really know what will happen, unlocked by Google’s efforts with TensorFlow really, which componentised the technology, what is the equivalent path for faster than real-time simulation?
If somebody can turn faster than real-time simulation into a new hammer, what nails could we hit?
There’s a new computer chip called the Cerebras Wafer-Scale Engine which is massive.
Chips are usually the size of a postage stamp, and you get faster computers by linking a bunch of them together. A different approach: the Cerebras is the size of an iPad.
The Cerebras has 1.2 trillion transistors, 20 times bigger than the world’s next largest chip, and 75 times bigger than the new Apple M1 chip. Singularity Hub has all the details: The Trillion-Transistor Chip That Just Left a Supercomputer in the Dust.
Anyway, this line from the announcement caught my eye:
Faster than real-time simulation.
I mean, that’s what physics does anyway, right? We can figure out the position of the planets in the solar system for a thousand years in the future without having to run the calculation for a thousand years.
And yet… something provocative about this.
One of the challenges with nuclear fission, the holy grail of clear energy generating, is the plasma going out of control, and the magnetic fields in the torus can’t be adjusted fast enough to contain it. The physics is too hard; nobody’s “solved” the plasma problem yet. But if the entire thing can be simulated from the bottom up, faster than realtime, you don’t need a model. You just run the simulation.
What are the civilian applications?
I can imagine a wearable device that continuously snapshots the world around you, runs the simulation in fast forward, and pre-emptively warns you about a mugging, or a bike going off course. Call it augmented apprehension.
Or how about intelligent fire extinguishers that simulate the fire in faster-than-realtime and dynamically direct the spray to uncannily effective spots.
I think it’s that uncanniness that draws me the most. Fluid dynamics and chaotic systems generally are weird and interesting. I think of the weird interference patterns you get in a pool of water if you get ripples to meet up in specific ways, or the strange behaviours of inverted pendulums that stand upright if you vibrate them at the right frequency. (Human skeletons are basically realtime adjusted inverted pendulums.)
So, with powerful simulation, could you figure out how to hit a mass of water with puffs of air so that it rises up and moves around the room, washing the windows; or robots with reed-thin jointed limbs that should never be able to hold themselves up, but with motors at each joint running at just the right vibration to keep the thing moving?
The general algorithm would seem to be:
…which is how dolphins swim and bumblebees fly.
Just as machine learning is getting into everything, and changing all software to the point that we don’t really know what will happen, unlocked by Google’s efforts with TensorFlow really, which componentised the technology, what is the equivalent path for faster than real-time simulation?
If somebody can turn faster than real-time simulation into a new hammer, what nails could we hit?