FEATHER: A Reconfigurable Accelerator with Data Reordering Support for Low-Cost On-Chip Dataflow Switching

Authors: Jianming Tong, Anirudh Itagi, Prasanth Chatarasi, Tushar Krishna | Affiliations: Georgia Tech; IBM Research | PDF: FEATHER_Reconfigurable_Accelerator_Dataflow_Switching_2024.pdf | arXiv: 2405.13170

一句话总结

FEATHER 用 NEST(2D PE + 行时分共享归约)和 BIRRD(蝶形 arbitrary reduce+reorder)实现每层 (dataflow, layout) co-switchRIR 在归约阶段隐式写出下一层 concordant layout,消除 Timeloop 忽略 layout 导致的 128× theory-practice 鸿沟,面积仅比固定 Eyeriss-like 基线 +6%

核心贡献

  1. 问题量化:dataflow 最优 ≠ 实践最优——bank port 冲突可致 128× 延迟差距;需 per-layer layout reorder(iActs online)
  2. NEST:local temporal reduction + 行间时分 spatial reduction,2D 阵列共享低成本归约网
  3. BIRRD + RIR:蝶形网络同时 arbitrary reduction 与 Arbitrary Reorder;reorder 延迟隐藏在 compute reduction 中(非 RAR/off-chip)
  4. Layoutloop:Timeloop + 物理 buffer 建模 + layout 评估 + dataflow-layout co-search

关键机制

Dataflow (TOPS)

Tiling / Ordering (stationarity) / Parallelism / Shape — 单层设计空间 O(10³⁶)。

Layout 术语

CHW W4H2C2 = inter-line 顺序 C→H→W + intra-line (4,2,2) 展平为 W→H→C。

Reorder 谱系

Fixed → Line Rotation → Transpose → Row Reorder → Arbitrary(FEATHER/BIRRD)

实现代价
Off-chipDRAM 往返 + CPU reorder
RAR (Medusa/MTIA/TPUv4)reorder 在 critical path
RIRoActs 归约时直接写入新 layout

实验摘要

对比结果
Layoutloop: vs NVDLA/Eyeriss/SIGMA延迟 1.27–2.89×,能效 1.3–6.43×
ZCU104: vs Gemmini / Xilinx DPU3.91× / 2.65× 归一化吞吐
vs Edge TPU4.56× geomean
面积 vs 固定 Eyeriss-like+6%
最优 (dataflow,layout) co-switch能效 +27–33%

与 wiki 交叉引用

Citations

[1] FEATHER_Reconfigurable_Accelerator_Dataflow_Switching_2024.pdf — Tong et al. (2024)