FCM: Towards Fine-Grained GPU Power Management
for Closed Source Mobile Games
Jiachen Song
1
, Xi Li
1,2
, Beilei Sun
1,2
, Zhinan Cheng
1
, Chao Wang
1,2
and Xuehai Zhou
1,2
1
Dept. of Computer Science and Technology, University of Science and Technology of China (USTC)
Hefei, 230027, China
2
Suzhou Institute for Advanced Study, USTC, Suzhou, 215123, China
{jchsong, sasbl, chinam}@mail.ustc.edu.cn, {llxx, cswang, xhzhou}@ustc.edu.cn
ABSTRACT
Contemporary mobile platforms employ embedded graphic
processing units (GPUs) for graphics-intensive games, and
dynamic voltage and frequency scaling (DVFS) policies are
used to save energy without sacrificing quality. However, cur-
rent GPU DVFS policies result in unnecessary power waste
due to defective workload estimations of embedded GPUs
during game play. In this paper, we propose the Frame-
Complexity Model (FCM), a fine-grained estimation of the
GPU workload in a game frame, to quantify the GPU work-
load with the real runtime demand for GPU computing re-
sources of a game frame. In FCM, three constituents of a
game frame (i.e., structure, textures and computation) are
quantified without modification of mobile games. Prelimi-
nary experiments show that, compared with the default pol-
icy, the FCM-directed GPU DVFS policy can reduce more
power consumption of games (11.3% to 25.8%) with good
Quality of Service (QoS).
Keywords
GPU Workload Model; DVFS; Mobile Games; Fine-grained;
Power Management
1. INTRODUCTION
3D mobile games compete for computing resources of both
CPUs and GPUs in mobile phones. Recently, GPUs have
been launched with increasing frequency to satisfy graphics-
intensive games [1]. Basically, high frequency leads to high
power consumption. Efficient power management of embed-
ded GPUs has become a critical problem [4, 7], especially
during game play [6].
DVFS is introduced as the main technique for power man-
agement of embedded GPUs on mobile platforms [8]. How-
ever, most of current policies [3, 6, 8], including the default
DVFS policy for embedded GPUs in smart phones, borrow
the CPU-like workload estimation, in which the workload is
estimated with the utilization of underlying GPUs in pre-
vious sampling intervals [7]. Unfortunately, in this histor-
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full cita-
tion on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-
publish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
GLSVLSI ’16, May 18-20, 2016, Boston, MA, USA
c
2016 ACM. ISBN 978-1-4503-4274-2/16/05. . . $15.00
DOI: http://dx.doi.org/10.1145/2902961.2902989
ical utilization-based technique, mistaken and lagged esti-
mations result in unnecessary power waste or performance
loss. These defects become even worse in highly interactive
mobile games: On the one hand, the GPU workload changes
harshly from frame to frame due to the unpredictable in-
put by players. On the other hand, the current workload is
estimated with the information in previous frames, which
may be outdated and wrong. Though more accurate mod-
els, such as PID [3] and LMS [2], are applied to improve
the estimation, they can not eliminate the inherent defects
of this historical utilization-based estimation. Besides, the
open-source-based estimation in [3] can hardly be applied
as most mobile games are closed source games. The game-
state-based method in [2] fails to provide a fine-grained es-
timation since the GPU workload changes a lot in the same
state (e.g., gaming). Thus, policies mentioned above lack
of feasibility in the design of efficient GPU DVFS policies
for closed-source mobile games. A different method of GPU
workload estimation is required.
We find that, in modern graphics-intensive mobile games,
a game frame mainly consists of three different kinds of G-
PU workloads, including the structural, the textural and the
computational workloads. Each of them can be well quanti-
fied by specific parameter in the data flow from the CPU to
the GPU. Besides, all the parameters can be easily traced
without modification of games in Android system. These
observations motivate us to build an accurate GPU work-
load model for runtime mobile games, which can be used
to direct the GPU DVFS for closed-source mobile games in
a fine-grained way. Totally, this paper makes the following
contributions:
• Find out quantifications (the number of vertices, the
size of textures and the amount of OpenGL|ES com-
mands) for main constituents of GPU workload (the
structural, textural and computational complexity) in
a game frame.
• Build the Frame-Complexity Model (FCM) for run-
time GPU workload estimation with found quantifica-
tions. Based on FCM, we propose the FCM-directed
GPU DVFS policy (FGDP).
• Realize FGDP on our smartphone-like platform. Ex-
periments show that FGDP saves more power (11.3%-
25.8%) than the default utilization-based policy with-
out obvious user-experience reduction.
The rest of this paper is organized as follows: In Section 2,
we describe the design of FCM and FGDP in detail, followed
by the experimental evaluation and analysis in Section 3. We
conclude our work and plan the future work in Section 4.