Solutions based on artificial intelligence represent a real gold mine, both for those who design and implement them, and for those who use them, especially in the corporate sector. The workload necessary for inference activities but, above all, for model training is increasing and puts even the most advanced configurations to the test.
Extropica startup led by former Alphabet X quantum researcher Guillaume Verdon, claims to be able to make AI accelerators which are many orders of magnitude faster and more energy efficient than traditional chips: CPU, GPU, TPU and FPGA. Confirming the interest surrounding the Extropic project, the company recently obtained 14.1 million dollars in financing.
A new era for computing with artificial intelligence and thermodynamic chips
In the introduction published on the Extropic website, Verdon and his team explain that it is essential to introduce a new paradigm for probabilistic computing, more distant from the solutions used to date. The prototypes of passive thermodynamic chips Extropic executes stochastic algorithms widely used in AI applications as physical, fast, and energy-efficient processes.
According to Verdon, the proposed solution would be much more in line with the nature of generative models compared to the use of traditional chips, which instead try to embrace probability and uncertainty in a rather unnatural way, introducing inefficiencies.
What Extropic is doing is integrating all the neural computing processes into one analog thermodynamic chip. Founder and CEO Verdon explains that the engineers at his startup have worked hard and continue to take what they have learned from quantum computing adapting it to a thermodynamic approach.
In the image that appeared on the Extropic website you can see two Josephson junctions, transistor components suitable for the types of complex nonlinear problems commonly faced by AI. The one shown is only a small module of the processor that will later be used in supercomputing systems.
Extropic joins other companies determined to rethink the role of traditional processors
In another article we talked about the historical transition from bits to qubits, with the future diffusion of quantum computers. Here, Extropic – like other companies – seem to be talking about a story that has already been written: current chips would not be able to keep up with the computational requirements and the growing dimensions of generative models.
He doesn’t think so NVIDIAwhich in March 2024 unveiled a nearly infinitely scalable up solution based on the new Blackwell architecture.
For Verdon and his team, however, the power of GPUs and other chips can certainly grow, but with it energy consumption and heat dissipation requirements will increase significantly. We therefore need an innovative solution, which in a future perspective “breaks” with tradition, although it has been used for 50 years now.
There are many challenges to face, however. Also because Extropic’s prototypes work at low temperatures to demonstrate the behavior of superconductors. The startup, however, is reportedly preparing devices capable of operating at room temperature. In this case, transistors will be used instead of Josephson junctions, sacrificing energy efficiency but opening up the possibility of using an expansion card.
The images published in the article are taken from the official Extropic website.