Gemmini: Enabling Systematic Deep-Learning Architecture Evaluation via\n Full-Stack Integration

DNN accelerators are often developed and evaluated in isolation without\nconsidering the cross-stack, system-level effects in real-world environments.\nThis makes it difficult to appreciate the impact of System-on-Chip (SoC)\nresource contention, OS overheads, and programming-stack inefficiencies on\noverall performance/energy-efficiency. To address this challenge, we present\nGemmini, an open-source*, full-stack DNN accelerator generator. Gemmini\ngenerates a wide design-space of efficient ASIC accelerators from a flexible\narchitectural template, together with flexible programming stacks and full SoCs\nwith shared resources that capture system-level effects. Gemmini-generated\naccelerators have also been fabricated, delivering up to three\norders-of-magnitude speedups over high-performance CPUs on various DNN\nbenchmarks.\n * https://github.com/ucb-bar/gemmini\n