Extropic is introducing a new paradigm for "thermodynamic computing," a new hardware platform that not only improves computing performance, but also fundamentally changes the way we think about the concept of "computing."
What?
The core concepts of thermodynamic calculations are:
Instead of suppressing the generation of entropy that is the origin of the world, we will achieve symbiosis with it, turning the ubiquitous thermal noise of nature into a force to advance computing. Its goal is to embed probabilistic generative AI algorithms directly into the physical world, so that the computational process fully conforms to the laws of thermodynamics, and achieves the ultimate efficiency of physics in terms of space, time and energy
The Extropic researchers claim to have pushed the boundaries of transistor technology to invent a new superconducting device that uses aluminum to restrain the electrons in the Cerfton diode at extremely low temperatures, which is the key to thermodynamic calculations
Why?
Bottlenecks in Traditional Computing Moore's Law has begun to slow down the increase in digital computing power due to the inability of transistor manufacturing processes to scale indefinitely. At the same time, physical constraints such as noise and energy consumption have put the capabilities of traditional computing architectures to a ceiling. Cellular network operations in living organisms, while inherently efficient, rely on random fluctuations in the origin of reactants rather than rigid number crunching. This inspired Extropic researchers to redefine the concept of "computer" and seek a new computing paradigm that blends with the real physical laws of the world.
Quantum computing has high hopes of pushing the limits of classical computing. But noise has always been its dead end. Extropic scientists from the field of quantum computing have discovered over the years that noise is not necessarily an obstacle, on the contrary, it can be the driving force behind computing!
They call it "thermodynamic calculations". Unlike numerical calculations, which encode data into 0s and 1s, thermodynamic calculations use the thermal motion of electrons in the presence of temperature to map data as a continuous fuzzy state. The optimization process of machine learning is like manipulating these quantum states to gradually reach the target state
This new computing paradigm not only improves energy efficiency, but more importantly, it is closer to the way nature operates
How?
Extropic is a practitioner of thermodynamic calculations. Its founder, who came from the field of quantum computing, saw that the path that relied on quantum mechanics still faced many difficulties in industrialization, so he abandoned quantum computing in favor of a new paradigm that uses noise to advance computing
How does Extropic implement thermodynamic calculations?
Its first-generation processor was made of aluminum and was superconducting when operating at extremely low temperatures. A variety of potential "superconducting neuron" designs are integrated on the chip, and the Josephson effect is used to introduce nonlinearity to form a high-dimensional non-Gaussian distribution. By adjusting neuronal parameters, a single device can generate a wide family of probability distributions. These passive superconducting chips consume energy only when measured or operated, making them extremely energy efficient
Extropic chips work similarly to Brownian motion:
Think of electrons as particles floating in liquid molecules, and the impact of the electrons on the surrounding atoms creates a thermal motion that spreads and drifts in the circuit. Extropic designed components in circuits that restrict the movement of electrons, such as coils and transistors, which act as spring bindings to Brownian particles. Under these constraints, the thermal noise of the electrons causes them to form a steady-state probability distribution
By adjusting the parameters of the circuit components, such as changing the coil inductance, transistor bias, etc., different probability distributions can be "programmed". This process is a bit like changing the stiffness of a spring to control the Brownian particles, except that in Extropic's devices, it is the electrons in the circuit components that play this role
Extropic is also developing a semiconductor version that can be operated at room temperature and can be produced at scale. At the software level, the company is building a layer that compiles probabilistic models to hardware, enabling oversized programs to be executed in layers on simulated or real thermodynamic accelerators
Extropic has raised $14.1 million in seed funding, attracting a number of well-known technology companies to accelerate the company's progress in thermodynamic computing
epilogue
This whole idea of leveraging analog circuitry to accelerate AI workloads is not new. IBM, Intel, and others have been trying for years to take advantage of the noise tolerance of neural networks to improve efficiency
The problem is that analog design is difficult, orders of magnitude harder than ordinary digital VLSI. Mixed-signal processing, nonlinear variations, device-level manipulation, and more
Extropic hasn't made any software simulations or benchmarks public, which is a violation of the first rule of hardware development for a hardware company that claims to have a revolutionary new architecture? Simulation precedes manufacturing
Of course, if it happens, it will completely disrupt the current form of computing