Energy-Quality Scalable Systems

Research Overview

Approximate computing has recently emerged as a promising approach to the energy-efficient computing of error-resilient applications where the quality of computation results can be traded for energy efficiency. While previous efforts to exploit this tradeoff have largely focused on computing subsystems (e.g., logic circuits and microarchitecture), in embedded systems where non-computing subsystems (e.g., sensors, actuators, user interfaces, network interfaces, etc.) consume a significant amount of energy, approximate computing alone can only exploit a narrow range of the energy-quality tradeoffs, leaving potentially greater tradeoffs in non-computing subsystems largely untapped.

The aim of our research is to develop a new design methodology to achieve full-system energy-quality scalability by answering the following questions:

What energy-quality scalability exists in subsystems of embedded systems?
How should the level of approximation be set at design-time and at run-time?
How might one derive an optimal energy-quality scalable system with less effort?

AxSerBus: Approximate Serial Bus

Mobile, wearable, and implantable devices integrate an increasing number and variety of sensors, such as microphones, image sensors, and accelerometers. These devices spend substantial amounts of time reading the sensors within them, incurring significant data transfer from the sensor to the processor. The high capacitance of sensor-to-processor interconnects, high data rates of modern sensors and frequent usage of applications that use sensors result in significant energy dissipation for sensory data transfer. The contribution of sensing to power consumption is only expected to increase as more and more high performance sensors are embedded in a single device.

AxSerBus is a quality-configurable approximate serial bus that exploits the locality of sensory data and the error resiliency of sensing applications to reduce energy dissipation. AxSerBus significantly reduces signal transitions by encoding the differences of sensory data in three encoding modes, depending on the magnitude of the differences. Compared with previous schemes, the proposed multi-level encoding results in more data being encoded using low-energy patterns.

Power saving by AxSerBus for image data transfer.

SAADI: Approximate Integer Divider

Division is an arithmetic operation crucial in signal processing. A hardware divider is a costly module in terms of latency and energy consumption due to the high complexity of division algorithms. For example, the integer divider instruction (IDIV) of the AMD 12h family has a latency of 9–17 cycles for a 16-bit division and 9–25 cycles for a 32-bit division, while the integer multiplier instruction (IMUL) for the same widths takes only three cycles.

SAADI is a novel approximate divider (Scalable Accuracy Approximate Divider), which is capable of exploiting the trade-off between accuracy and latency, thus energy. It finds the approximate reciprocal of the divisor, which is multiplied by the dividend to obtain the division result. The iterative reciprocal approximation process, where the accuracy gradually increases as more number of iterations are performed, enables the accuracy-energy trade-off. The number of iterations can be dynamically adjusted by the application for energy-quality scaling–the application can either obtain high-accuracy division results at the cost of a higher latency and energy consumption or reduce latency and improve energy efficiency by sacrificing accuracy.

Research Outcomes

Awards

Full-System Quality-Configurable Approximate Computing
Younghyun Kim
Best Presentation Award @ KOCSEA Technical Symposium, 2018
AxSerBus: A Quality-Configurable Approximate Serial Bus for Energy-Efficient Sensing
Setareh Behroozi, Jingjie Li, Vijay Raghunathan, Anand Raghunathan, Younghyun Kim
Low-Power Design Contest Award @ ISLPED (International Symposium on Low Power Electronics and Design), 2018

Publications

SAADI: A Scalable Accuracy Approximate Divider for Dynamic Energy-Quality Scaling
Setareh Behroozi, Jingjie Li, Jackson Melchert, Younghyun Kim
ASP-DAC (Asia South Pacific Design Automation Conference), 2019
A Quality-Configurable Approximate Serial Bus for Energy-Efficient Sensory Data Transfer
Setareh Behroozi, Vijay Raghunathan, Anand Raghunathan, Younghyun Kim
IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2018
AxSerBus: A Quality-Configurable Approximate Serial Bus for Energy-Efficient Sensing
Younghyun Kim, Setareh Behroozi, Vijay Raghunathan, Anand Raghunathan
ISLPED (International Symposium on Low Power Electronics and Design), Taipei, Taiwan, 2017

Invited Talks

Full-System Quality-Configurable Approximate Computing
Korean Computer Scientists and Engineers Association in America (KOCSEA) Technical Symposium, November 2018
Beyond Approximate Computing: Enabling Full-System Energy-Quality Scalability
SK hynix America, October 2018