This post is also available in:
עברית (Hebrew)
Some computing problems become difficult not because any individual calculation is particularly complicated, but because there are simply too many possible combinations to test. Designing a computer chip is one example: engineers must decide how thousands of components should be connected, where those connections should run and which wiring layers should carry them, all while working within limited physical space.
Researchers have developed a new spintronic Ising machine designed to tackle these combinatorial optimization problems using magnetic states rather than relying entirely on conventional computing.
The machines represent possible choices as “spins”, variables that can occupy one of two states. By repeatedly changing those states and allowing them to interact, the system searches for configurations that provide strong solutions to a given optimization problem.
According to TechXplore, the new design combines conventional CMOS electronics with magnetoresistive random-access memory (MRAM) containing 96,000 spin elements. Each represents either +1 or -1 through its magnetic state.
Electrical pulses control whether those magnetic elements switch between states. By adjusting the duration of each pulse, researchers can set the probability of a switch anywhere between 0% and 100%.
Those updates take just 0.3 to 1 nanosecond, while consuming less than 40 femtojoules per spin.
Researchers demonstrated the machine on practical chip-design tasks, including finding efficient wiring routes between components and determining which wiring layers should carry individual connections. It was also evaluated using Max-cut optimization benchmarks, reaching a reported system efficiency of 1.92 × 10⁵ solutions per second per watt.
Although this research focuses on chip design and mathematical optimization, the underlying challenge has parallels elsewhere in computing, including AI.
Every AI system depends on the dataset used to train it, and those databases can contain incorrect labels, duplicated records, missing information and anomalous samples. Finding the best way to identify, clean and organize enormous datasets is itself an optimization challenge.
Specialized AI-data technologies can already approach this problem by automatically finding labeling errors and duplicates, detecting suspicious outliers, filling data gaps or generating synthetic training data where real examples are insufficient. This is a different application from the spintronic Ising machine described in the research, but it reflects the same broader need: extracting better results from increasingly complicated combinations of information.
For defense AI, data quality can be particularly consequential. Systems developed for autonomous platforms, intelligence analysis, threat recognition or computer vision are only as reliable as the information used to train and validate them.
That creates an opportunity for companies whose technology improves AI datasets even if they do not currently consider themselves defense companies. A tool built to clean and optimize commercial training data could become infrastructure supporting many different military AI applications.
This challenge is relevant to INNOFENSE, the innovation program operated by iHLS in cooperation with the Israeli Ministry of Defense and DDR&D (MAFAT). Developers working on dataset cleaning, automatic labeling validation, anomaly detection, missing-data reconstruction or synthetic data can apply with technologies that could be adapted and demonstrated through a POC with the defense establishment.
The individual AI model often gets the attention. But before any model can perform reliably, someone has to make sure the data feeding it is worth learning from.
The research was published here.
Are you working on a technology that might relate to optimization and think it might be able to apply to AI-based DB optimization? Apply to INNOFENSE now!
