Highlights:
- A significant obstacle in the quest to create a commercially viable quantum chip is the high susceptibility of these processors to computing errors, commonly known as noise.
- Should the report about Extropic developing an LLM-optimized processor prove accurate, the company will likely encounter competition from Nvidia Corp., who recently unveiled the H200, a specialized data center processor designed explicitly for running LLMs.
Extropic, a hardware startup guided by former individuals from Alphabet Inc.’s quantum computing research team, has officially disclosed securing USD 14.1 million in seed funding. Kindred Ventures led the investment. HOF Capital, Marque VC, Valor Equity Partners, Julian Capital, OSS Capital, and Weekend Fund contributed to this funding round. The funding round for Extropic also witnessed the involvement of over a dozen other supporters, which included executives from companies such as Adobe Inc., Shopify Inc., and various venture-backed artificial intelligence startups.
Founded just last year, Extropic is led by Chief Executive Officer Guillaume Verdon, who brings his prior experience of heading a quantum computing team at Alphabet’s X research unit. Extropic’s Chief Technology Officer, Trevor McCourt, also brings valuable experience as a former researcher at the renowned search giant. During their tenure at Alphabet, Verdon and McCourt spearheaded the development of a TensorFlow library designed for running AI models on quantum computing chips.
Extropic reportedly constructs a chip optimized explicitly for running large language models or LLMs. In a recent blog post, Verdon elucidated the company’s technology, portraying it as a “novel full-stack paradigm of physics-based computing” and explaining that it harnesses “out-of-equilibrium thermodynamics.” This suggests that Extropic’s chip design integrates principles from non-equilibrium thermodynamics, a burgeoning field of physics dedicated to examining phenomena like chemical reactions.
Verdon’s blog post says the company’s product is not a quantum computing chip. He penned, “As the timelines to scalability for quantum physics-based computers grew endlessly longer and longer, many of our team sought a different path to practical physics-based computing.”
A significant obstacle in the quest to create a commercially viable quantum chip is the high susceptibility of these processors to computing errors, commonly known as noise.
The prevalence of errors in quantum processors poses a significant challenge, rendering it nearly impossible to execute intricate calculations with the required reliability. Extropic aims to construct a system where “noise is an asset rather than a liability,” challenging the conventional view of noise as a liability in computing systems.
Extropic has yet to disclose more detailed technical information about its technology. Nevertheless, Verdon has disclosed that one of Extropic’s objectives is to minimize the electricity consumption required to operate AI models. He also hinted that the company’s technology will automate specific coding tasks, expressing that “one could imagine a computer which, instead of being imperatively programmed, naturally finds a way to program itself to learn representations of the world.”
The fact that Extropic’s technology is designed to run AI models suggests that there may be additional details about its functionality and capabilities.
Neural networks process data through matrix multiplications, a mathematical calculation set on matrices. Matrices are mathematical structures that consist of numbers arranged in rows and columns, resembling a spreadsheet. Almost all AI-optimized chips incorporate circuits designed explicitly for matrix multiplications, as these operations play a crucial role in the data processing performed by AI models.
AI chips generally integrate a substantial amount of high-speed memory. This is because AI models often require the swift movement of data to and from memory during the processing phase. A neural network can generate results for users more rapidly when the data efficiently completes the round trip, moving swiftly to and from memory during processing.
Should the report about Extropic developing an LLM-optimized processor prove accurate, the company will likely encounter competition from Nvidia Corp., which has recently unveiled the H200, a specialized data center processor designed explicitly for running LLMs. Distinguishing itself from Nvidia’s prior flagship graphics card, the H200 boasts twice the onboard memory.