Work

FPRaker: Exploiting Fine-grain Sparsity to Accelerate Neural Network Training

Year: 2020
Type: dissertation
Abstract: sis presents FPRaker, a processing element for composing training accelerators. Training manipulates floating-point data and multiply-accumulate (MAC) operations constitute the bulk of its computations. FPRaker boosts performance and energy-efficiency by skipping ineffectual computations during training. FPRaker processes the operands’ significand of each MAC as a series of signed-powers-of-two, or terms. This exposes ineffectual work that can be skipped: encoded values have few terms and some can be discarded as they would fall outside the accumulator's precision. Over 9 studied networks, FPRaker is 1.5x faster and 1.4x more energy-efficient compared to a baseline accelerator with conventional TensorCore-like tiles under iso-compute-area constraints. We demonstrate that FPRaker delivers additional benefits when training incorporates pruning, quantization and methods that use a different accumulator precision per layer. Finally, we propose a memory compression technique for exponents of floating-point values that exploits the narrow value distribution during training using base-delta compression reducing off-chip memory bandwidth.
Authors: Mohamed M. Awad, Omar Amin
Language: English
Open Access status: green
OpenAlex ID: W3207828344
Cites: 85
Cited by:
Related to: 10