CAREER: Cross-Layer Integration of Stochastic Processing Ensembles for Versatile Intelligent Computing

The extraordinary advances of Artificial Intelligence (AI) have been demonstrated in various aspects of our lives, from healthcare to transportation. However, the ever-growing capability of AI is accompanied by an exponential increase in computational complexity, leading to unprecedented energy consumption and ever-growing demand for computing/memory resources.

As a result, developing efficient and scalable intelligent hardware systems is critical for (1) addressing the environmental and sustainability implications of AI computation and (2) enabling broad adoption of AI to improve the productivity and economic competitiveness of our society. One promising direction for developing efficient AI computing systems is to exploit stochasticity (randomness) to process the various statistical learning models.

However, the hardware implementation of truly-random functionality on standard complementary metal-oxide-semiconductor (CMOS) components lacks efficiency due to significant area and power overheads. Moreover, the conventional approach of incorporating stochasticity for brain-inspired computing focuses on cell-level emulations of neurons and synapses, facing significant challenges in scalability.

In this project, a novel system architecture with emerging devices/circuits will be developed to create intelligent computing systems based on ensembles of stochastic processing elements. The knowledge on designing scalable and robust AI computing systems created from this project will be disseminated to a broader community of our society through a multi-level portfolio of educational efforts.

Such educational efforts are designed to advance AI literacy among a population with diverse backgrounds and cultivate a strong future workforce with cross-disciplinary expertise in AI and microelectronics.

To accelerate various critical AI operations from matrix-vector multiplications to softmax in the attention layer, device-architecture-system cross-layer co-optimization will be conducted to integrate spintronics with intrinsic stochasticity and deterministic silicon components into energy-efficient non-von Neumann computing systems. Next, ensemble-based hardware-aware computational models will be developed to enable such novel hardware fabrics for large-scale learning tasks.

Furthermore, this project will design a unified and versatile hardware platform that supports a broad range of AI applications, including deep neural networks and emerging neuro-symbolic models. Such development will tackle the long-standing challenges of scalability and flexibility associated with stochastic computation and create a new venue for developing next-generation versatile intelligent computing systems.

This project is jointly funded by the Software and Hardware Foundations (SHF) program in the Computing and Communication Foundations (CCF) division and the Established Program to Stimulate Competitive Research (EPSCoR).

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.