IEEE Consumer Electronics Magazine - May 2018 - 78

Average Score

A detailed evaluation of HoloLens's
computing performance when running
5
HoloLens
our application is also conducted. As
Laptop
shown in Figure 8, before we started running our application (before t 1), the CPU
4
use is approximately 8-20%, HPU
engine #0 use is approximately 5%, and
3
memory is approximately 40% (i.e., 800
megabytes out of two gigabytes). Beginning at time t 1, we started using gaze and
2
gesture commands to open our application. Then both the CPU and HPU were
1
in heavy use, and memory use began rising. After we completely opened our
application at t 3, the CPU use dropped,
0
Natural
Visual Combination System
Flexibility
General
though approximately 10% was still used
Interaction
Quality
with Haptic Response
Rating
by our application. HPU use varies based
Devices
Time
on the visualized data size in the user's
Rating Aspects
view and occasionally has slight glitches
due to heavy use. The memory stays at
FIGURE 7. The user study results showing the scores of HoloLens and a traditional laptop
in natural interaction, visual quality, combination with haptic devices, system response 80% (1.6 gigabytes), which means ap
proximately 800 megabytes are occupied
time, flexibility, and general rating.
by our application.
Because HoloLens's performances
are
also
influenced
by
data
sizes, we conduct analysis with
obtains average scores of 3.4, 3.0, 1.6, 4.2, 2.3, and 3.1, with
various data sets. All values are recorded when the whole augstandard deviations of 1.2, 0.8, 0.4, 1.0, 1.2, and 0.8. HoloLens
mented city contents are within HoloLens's field of view.
wins all aspects except system response time. Finally, each user
Table 4 lists five situations: no city model (before opening our
is asked to choose one method as the preferred method. Eight
application), pure city model (75 megabytes without city data
users chose HoloLens, and the other four users, who selected the
sets), and with city data sets places of interest (65 kilobytes),
laptop, mainly complained about the HoloLens system
youth services (326 kilobytes), and heritage inventory (1,285
response time and gesture-recognition accuracy. Therefore, we
kilobytes), separately. The Toronto 3-D city model is comcan conclude that HoloLens holds several advantages in visualposed of .fbx files, and data sets are in XML file format, as
izing city data over the traditional method, but it requires
previously mentioned. From Table 4, we can see that memory
improvements in computing power and gesture interaction.
use rises with city data size increase, i.e., visualizing a larger
Compared with HoloLens, the VR device has large limitadata set requires more memory space. When viewing the
tions on user movements because it isolates users from the
whole augmented city model, HPU use rises above 90%.
real surrounding environment. As a result, though it can also
Compared with visualizing a pure city model, it goes up a litdisplay the city model with various sizes and angles, users
tle when showing extra data set information on the model.
are not able to walk around the city model to look at different
SoC and system power consumption mainly change with HPU
parts naturally, considering the dangers of stumbling or hituse in this condition. CPU use rarely changes with different
ting a wall. In addition, in a meeting scenario, participants
data size and is actually just around 3% by our application.
wearing VR devices cannot use natural body or eye commuThe rendering performance, measured by frames per second
nication among them, whereas AR users can achieve it with
(fps), also changes under different conditions. When there is
the AR device's see-through lenses.
no city model, whether the main menu is displayed or not, the
frame rate is approximately 60 fps; after loading a pure city
COMPUTING POWER OF HOLOLENS
model, it changes to 3.8-3.9 fps, though it rises to 25 fps
Because HoloLens itself is a complete AR system, integrating
when looking at the city-data-type menu; with additional city
all components into one head-mounted device, it processes
data sets, there are only slight differences (with places of
all tasks with its own CPU and HPU, as mentioned previousinterest set, 3.8-4.1 fps, whereas with the other two sets, 3.8-
ly. Its CPU is a 14-nm Intel 32-bit system-on-a-chip (SoC)
4.2 fps). The frame rate is not high, because once we select
with one gigabyte of RAM, and its custom-designed HPU is
city data to visualize, the graphics processing unit use is high,
a TSMC-fabricated 28-nm coprocessor with 24 Tensilica digand the augmented contents do not change a lot.
ital signal processor cores arranged in 12 clusters occupying
Based on the evaluation results, we may improve our
another one gigabyte of double-data-rate RAM and eight
implementation with the following aspects in the future.
megabytes of static RAM [18]. In addition, the storage size is
First, current model sizes are small compared with the total
64 gigabytes.
78 IEEE Consumer Electronics Magazine

MAy 2018

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May 2018