Buffets: An Efficient and Composable Storage Idiom for Explicit Decoupled Data Orchestration

Scope

Accelerators often implement custom on-chip buffering logic that is difficult to reuse. Buffets define a small storage abstraction that can be composed across accelerator designs.

Abstract

Custom buffers consume design effort and bind storage behavior to one accelerator. Buffets provide decoupled fills and accesses, fine-grained synchronization, hierarchical composition, and multicast behind a reusable interface. An RTL implementation adds roughly 2% control overhead to an 8 KB RAM. Across several workloads, the design improves energy-delay product compared with DMA-managed double buffering and caches.