Does Arithmetic Logic Dominate Data Movement ? A Systematic Comparison of Energy-Efficiency for FFT Accelerators

Tung Hoang; Amirali Shambayati; Henry Hoffmann; Andrew Chien. 30 March, 2015.
Communicated by Andrew Chien.


Data format and data type selections for hardware accelerators design are important since it comes at the energy cost of arithmetic logic. However, these selections can affect the energy cost of data movement, as a large fraction of system energy, due to many memory-limited executions that are common on accelerator-centric heterogeneous systems. In this paper, we perform a systematic comparison to study the energy cost of varying data formats and data types w.r.t. arithmetic logic and data movement for heterogeneous systems in which both compute-intensive (FFT accelerator) and data-intensive accelerators (DLT accelerator) are added. We explore evaluation for a wide range of design processes (e.g., 32nm bulk-CMOS and projected 7nm FinFET) and memory systems (e.g., DDR3 and HMC).

Our results show that when varying data formats, the energy costs of using floating point over fixed point are 5.3% (DDR3), 6.2% (HMC) for core and 0.8% (DDR3), 1.5% (HMC) for system in 32nm process. These energy costs are even decreasing to only 0.2% and 0.01% for core and system in 7nm FinFET process respectively with DDR3 memory (slightly increasing with HMC). Furthermore, we identify that the core and system energy of systems using fixed point, 16-bit, FFT accelerator is nearly half of using 32-bit which show the high proportion of system energy on the amount of moving data which is dependent on the selection of data type.

Original Document

The original document is available in PDF (uploaded 30 March, 2015 by Andrew Chien).