Artificial intelligence can now sharpen blurry photos, turn messy voice recordings into clear text, and even write emails in seconds.
These tools feel almost magical, but behind the scenes they require enormous amounts of energy and computing power.
Modern AI systems are extremely expensive to run. Training a large AI model such as GPT-3 can cost more than $10 million and consume huge amounts of electricity and water.
Even everyday AI tasks continue to rely on large energy-hungry computer systems.
One major reason for this problem is that today’s computers still follow a design created more than 75 years ago.
In traditional computers, memory and processing are separated. This means data constantly travels back and forth between storage and the processor.
For AI systems that handle massive amounts of information, this movement creates a serious bottleneck that wastes both time and energy.
Professor Ang Kah Wee from the National University of Singapore says that AI tasks are heavily focused on memory use. According to him, the biggest issue is not the computing itself, but the constant movement of data.
To solve this problem, researchers are developing a new technology called “compute-in-memory,” or CIM.
Instead of separating memory and processing, CIM combines them in the same place. This could make AI systems much faster and far more energy efficient.
The new system developed by Prof Ang’s team uses devices called memristors. Unlike ordinary transistors used in current computers, memristors can both store and process information. They also continue to remember data even when the power is turned off.
Researchers often compare memristors to the human brain because they work in a similar way. Instead of relying only on electrons, they use ions to carry information and can switch states very quickly while using little energy.
The team built a fully working CIM system using a 32-by-32 array of memristors made from an ultra-thin material called hafnium diselenide. To keep the system stable, each memristor was paired with a silicon-based selector that controls electrical flow and prevents signal interference.
The researchers also designed a new sensing method that measures tiny timing changes in electrical signals instead of using traditional converters, which are large and consume more power. This approach cut energy use by more than half.
Another advantage is that the system naturally performs some functions needed for neural networks, similar to how brain cells activate and communicate. Because these functions are built directly into the hardware, the system avoids extra processing steps and becomes even more efficient.
In testing, the memristor system completed image pattern recognition tasks with 97.5% accuracy, similar to conventional digital systems but using only a fraction of the energy.
The researchers believe this technology could become especially important for edge AI devices, autonomous systems, and future brain-inspired computers where low power consumption is critical.
Since the design is compatible with existing silicon manufacturing methods, it may also be practical for large-scale production in the future.
