2025 Theses Doctoral

Leveraging Design-Time Abstractions for Accelerator Design and Integration

Barker, Martha

With the end of Dennard scaling and the slowdown of Moore’s law, the computing industry has turned to heterogeneous architectures to improve performance and energy efficiency. These systems combine general-purpose processors with domain-specific accelerators, delivering significant performance-per-watt improvement by executing critical compute kernels in hardware. However, designing and programming accelerators remains a specialized and complex task, requiring both domain-specific knowledge and hardware design expertise.

This dissertation supports the thesis that design-time abstractions are essential to effective accelerator design and integration, and their careful selection can substantially simplify the implementation of key features. To support this claim, this dissertation presents three contributions that leverage abstractions to improve the design and usability of accelerators.

First, it introduces a synthesized, hardware-only garbage collector for FPGA accelerators. The collector operates concurrently with the accelerator, exploiting idle memory cycles and incurring negligible performance overhead under an eager collection policy. Second, it presents a method for automatic partitioning and control of fine-grained power domains in dataflow circuits. By using activity signals inherent to the dataflow abstraction, the method partitions logic into sleepable and always-on power domains and synthesizes control signals with provable correctness guarantees. Finally, it proposes Accel, a novel MLIR dialect for compiler-directed task dispatch in chiplet-based accelerators. Our dialect captures task graph structure and scheduling constraints, enabling a pull-based runtime execution model that improves locality and reduces synchronization overhead. Together, these contributions show that thoughtfully chosen abstractions can enhance productivity and enable features that would otherwise be difficult to implement.