UMass Amherst Engineers Develop Hardware-Software Approach to Make Edge AI More Efficient

Researchers at the University of Massachusetts Amherst are working on a hardware-software co-design approach aimed at making artificial intelligence more efficient on edge devices. The research addresses a key challenge in edge computing: delivering advanced AI capabilities while operating within strict limits on power, memory and processing resources.

Edge devices increasingly need to process AI workloads locally rather than continuously sending information to cloud servers. This can reduce latency and network traffic, but conventional computing architectures can consume significant energy because data has to move repeatedly between memory and processing units.

The UMass Amherst research focuses on redesigning both the AI algorithm and the underlying hardware, allowing the two components to work together more efficiently.

Redesigning AI Algorithms and Hardware Together

A major challenge with current AI systems is the separation between memory and processing. Moving data between these components can become an energy-intensive operation, particularly for edge devices with limited power budgets.

Compute-in-memory architectures offer one way to address this issue by allowing data to be stored and processed in the same location. However, conventional compute-in-memory designs can struggle with the complex mathematical operations required by modern AI models.

The researchers therefore explored changes to the AI model itself alongside modifications to the hardware architecture. One approach involves simplifying state-space models by replacing complex-number calculations with real-number operations. This allows individual memory cells to represent data more directly and reduces the amount of circuitry required for computation.

Improving Efficiency for Edge Devices

The combined hardware and algorithm approach is designed to reduce unnecessary computation and limit the movement of data within a chip. This can potentially lower energy consumption while improving processing performance.

Such improvements could be particularly valuable for edge applications where devices need to respond quickly without relying on a remote data centre. Examples include autonomous systems, robotics, wearable electronics and other connected devices that need to process information locally.

UMass Amherst research into edge AI also includes specialised hardware designed to bring intelligence closer to where data is generated. Researchers have been investigating technologies such as memristors and neuromorphic computing to reduce data movement and improve the energy efficiency of AI workloads.

Memristors Offer Another Route to Efficient AI

The university’s work on memristor-based computing provides another example of how hardware can be redesigned for AI workloads. Memristors can store information while also performing computation, reducing the need to move data between separate memory and processing components.

Researchers led by Qiangfei Xia have demonstrated memristive systems for in-memory and neuromorphic computing, with potential applications in low-power AI and intelligent sensing.

The university has also explored systems that process only relevant information rather than continuously analysing every data point. In one proof-of-concept sensing system, a memristor-based architecture focused processing on pixels containing useful signals, helping reduce unnecessary computation and improve energy efficiency.

Edge AI Moves Towards More Efficient Computing

The research highlights a broader shift in AI hardware development. Rather than relying solely on more powerful processors, researchers are increasingly looking at ways to redesign algorithms, memory architectures and processing hardware as a single system.

This approach could become increasingly important as AI moves into smaller devices with limited power and thermal budgets. By reducing data movement and unnecessary calculations, hardware-software co-design could help bring more capable AI models to the edge while maintaining practical energy requirements.

For the semiconductor industry, the development points towards a growing market for specialised AI architectures that combine algorithmic optimisation with purpose-built hardware. As edge AI adoption expands, such approaches could help enable faster, more energy-efficient intelligent devices.