Browse Prior Art Database

Method for execution-driven estimation of processor power consumption

IP.com Disclosure Number: IPCOM000124293D
Publication Date: 2005-Apr-14
Document File: 5 page(s) / 22K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a method for execution-driven estimation of processor power consumption. Benefits include improved functionality and improved performance.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 29% of the total text.

Method for execution-driven estimation of processor power consumption

Disclosed is a method for execution-driven estimation of processor power consumption. Benefits include improved functionality and improved performance.

Background

      Conventional design space exploration of system-on-a-chip architectures is performed to achieve higher performance with the smallest possible die size. The architecture is optimized for power consumption when the targeted performance has been achieved.

      Power consumption has become a major design constraint. As a result, a requirement exists to integrate the power model into the simulation infrastructure.

      The conventional solution for power estimation is based on analyzing small isolated kernels with microarchitecture-level activity factors.

General description

      The disclosed method is a power model for conventional and future processors based on instructions and events. This method is based on the fact that power consumption can be computed as the superposition of orthogonal effects.

              The power required for an instruction is the sum of the base power for an instruction, which includes the following factors:

•             Leakage power

•             Decoding

•             Execution

•             Retirement

•             Operand value

•             Register name switching

•             Stalls

•             Cache accesses

              To isolate the contribution of different orthogonal effects, short programs (called stimuli) can be written that estimate the incremental impact of each effect in isolation. Stimuli are small directed assembly programs that exercise specific functionality repeatedly in a loop.

 The energy cost of effects that cannot occur in isolation cannot be estimated using stimuli. However, they can be estimated through linear equations using multiple stimuli.

Advantages

              The disclosed method provides advantages, including:
•             Improved functionality due to integrating power estimation into an execution-driven simulation infrastructure

•             Improved functionality due to enabling the estimation of power consumption over time

•             Improved performance due to running real applications at a speed higher than a million instructions per second

Detailed description

              The disclosed method can be partitioned into three successive phases:

•             Characterization

•             Execution-driven estimation for real workloads (full applications) running on a simulator

•             Validation

Characterization

              Stimuli isolate the impact of a single effect, such as an instruction opcode, a register update, or a stall. However, some effects cannot be characterized directly by a stimulus. In this case, stimuli are linearly combined to isolate the impact of a single effect.

              Each stimulus is run on a system that enables processor power measurement or on microarchitecture or RTL models with power modeling ca...