Bit Based Approximation for Approx-NoC: A Data Approximation Framework for Network-On-Chip Architectures
Abstract
The dawn of the big data era has led to the inception of new and creative compute paradigms that utilize heterogeneity, specialization, processor-in-memory and approximation due to the high demand for memory bandwidth and power. Relaxing the constraints of applications has led to approximate computing being put forth as a feasible solution for high performance computation. The latest fad such as machine learning, video/image processing, data analytics, neural networks and other data intensive applications have heightened the possibility of using approximate computing as a feasible solution as these applications allow imprecise output within a specific error range.
This work presents Bit Based Approx-NoC, a hardware data approximation framework with a lightweight bit-based approximation technique for high performance NoCs. Bit-Based Approx-NoC facilitates approximate matching of data patterns, within a controllable error range, to compress them thereby reducing the data movement across the chip. The proposed work exploits the entropy between data words in order to increase their inherent compressibility. Evaluations in this work show on average 5% latency reduction and 14% throughput improvement compared to the state of the art NoC compression mechanisms.
Citation
Siravara, Divya Ravi (2019). Bit Based Approximation for Approx-NoC: A Data Approximation Framework for Network-On-Chip Architectures. Master's thesis, Texas A & M University. Available electronically from https : / /hdl .handle .net /1969 .1 /185084.