Matrix multiplication is one of the quiet engines behind modern AI. A new theoretical advance has shown that large matrices can be multiplied more efficiently than previously known, a result that could matter for chatbots, image systems, video synthesis, and other computing tasks that depend on this core operation.
Why matrix multiplication matters
Matrix multiplication is the process of combining rectangular arrays of numbers. It is central to AI systems such as ChatGPT, speech and image recognition models, AI image generators, chatbots from major vendors, and video synthesis models like Sora.
The same math also appears beyond AI. The source article points to image processing and data compression as examples of areas where matrix math is important. Because the operation is so common, even a small improvement can have wide implications for computation and power use.
GPUs are especially useful for this kind of work because they can handle many calculations at the same time. They split a large matrix problem into smaller parts and process those parts concurrently. That makes the algorithm used for matrix multiplication a major target for efficiency gains.
The new advance lowers omega
The recent work comes from Ran Duan and Renfei Zhou of Tsinghua University, Hongxun Wu of the University of California, Berkeley, and, in a second paper, Virginia Vassilevska Williams, Yinzhan Xu, and Zixuan Xu of the Massachusetts Institute of Technology.
The research focuses on the complexity exponent known as omega, or ω. In plain terms, omega describes how the number of required operations grows as matrix size increases. The closer omega gets to 2, the closer the method gets to the theoretical best possible efficiency.
The traditional method for multiplying two n-by-n matrices requires n-cubed separate multiplications. For a smaller example, multiplying two 3×3 matrices in the traditional way requires 27 multiplications. The newer line of work reduces how quickly the workload grows as matrices get larger.
Earlier techniques built on the “laser method” introduced Volker Strassen in 1986 and improved by Alman and Williams in 2020. The latest reported result brings the upper bound for omega down to approximately 2.371552.
That number may look like a tiny movement on paper. But the source says the reduction from the 2020 record value is 0.0013076, and describes it as the most substantial progress in the field since 2010.
What changed in the method
The breakthrough came from identifying a previously unknown inefficiency in approaches related to the laser method. The source describes the issue conceptually as a “hidden loss,” where useful blocks of data may have been underused or thrown away during the calculation framework.
In matrix multiplication, blocks are smaller pieces of a larger matrix. Breaking a large matrix into blocks makes the problem easier to manage, especially when many operations can happen in parallel.
The newer approach analyzes these blocks more carefully. It can involve labeling blocks so researchers can better decide how they should be combined. By using the available structure more effectively, the researchers were able to reduce the known upper bound for omega.
The timeline also matters. In late 2023, based on work updated in November, Ran Duan, Hongxun Wu, and Renfei Zhou revealed a method that set a new upper bound for omega at approximately 2.37188. Then, with findings presented in January 2024, Duan, Wu, and Zhou published further optimizations that lowered the figure to 2.371552.
“This is a major technical breakthrough,” said William Kuszmaul, a theoretical computer scientist at Harvard University, as quoted by Quanta Magazine. “It is the biggest improvement in matrix multiplication we’ve seen in more than a decade.”
What it could mean for AI
The practical promise is straightforward: if matrix multiplication takes fewer computational steps, AI workloads could eventually become faster or more efficient. That could affect training times, execution of AI tasks, and the ability to work with more complex models.
The potential benefits described in the source include:
- faster training times for AI models;
- more efficient execution of tasks that rely heavily on matrix multiplication;
- lower computational power requirements;
- lower energy consumption;
- reduced environmental impact from AI workloads.
Those outcomes are not automatic. The source makes clear that the exact effect depends on the architecture of a given AI system and how much of its workload depends on matrix multiplication. Algorithmic improvements often need hardware optimizations before their full value appears in real systems.
This also distinguishes the new research from the October 2022 work involving AlphaTensor, a Google DeepMind AI model. AlphaTensor focused on practical algorithmic improvements for specific matrix sizes, including 4×4 matrices. The newer work is more theoretical and aims at a broader gain across matrix sizes by lowering omega.
A breakthrough, but not the end
The result is important because it improves a foundational piece of computing rather than a single product feature. Matrix multiplication sits underneath many systems, so advances in the math can accumulate over time.
At the same time, the current approach has limits. The source notes that researchers expect further progress, but that deeper understanding is still needed before better algorithms can emerge.
“People are still in the very early stages of understanding this age-old problem.”
For AI, the main takeaway is cautious but significant. This is not simply a switch that makes every model faster immediately. It is a mathematical step toward more efficient computation, and if future algorithms and hardware can make use of it, matrix multiplication may become one of the places where AI gains speed without simply demanding more power.