Product Brief

Autonomous Driving

Intel and Mobileye Autonomous

Driving Solutions
Together, Mobileye and Intel deliver scalable and versatile solutions using
purpose-built software and efficient, powerful compute that will make
autonomous driving an economic reality.

Total Technology and Resources

Making the self-driving car a reality requires an unprecedented integration of
technology and expertise. Together, Mobileye and Intel provide comprehensive,
scalable solutions to enable autonomous driving engineered for safety and
affordability. With a combination of industry-leading computer vision, unique
algorithms, crowdsourced mapping, efficient driving policy, a mathematical model
designed for safety, and high-performing, low-power system designs, we deliver a
versatile road map that scales to millions, not thousands, of vehicles.

Purpose-built software and hardware

Mobileye’s portfolio of innovative software across all pillars of autonomous
driving (the “eyes”) complements Intel’s high-performance computing and
connectivity expertise (the “brains”). This winning package supports complex and
computationally intense vision processing with low power consumption, creating
powerful and smart autonomous driving solutions from bumper to cloud.
We start with revolutionary software from Mobileye that includes:
• Low-cost, 360-degree surround-view computer vision
• Crowdsourced Road Experience Management™ (REM™) mapping and localization
• Sensor fusion, integrating surround vision and mapping with a multimodal sensor system
• Efficient, semantic-based artificial intelligence (AI) for driving policy (decision-making)
• A formal safety software layer via our Responsibility Sensitive Safety (RSS) model

This advanced software calls for a powerful compute platform. We use a

combination of proven Mobileye EyeQ® SoCs for computer vision, localization,
sensor fusion, trajectory planning, and our RSS model designed for safety, as well
as Intel® processors for translation of the planned trajectory into commands that
are issued to the vehicle’s actuation system.

Production-ready solutions
With a powerful combination of software and hardware, all developed under one
roof, we can deliver solutions that scale to all levels of autonomous driving. Our
customizable, production-ready solutions deliver an optimal mix of compute
power, energy consumption, and cost.
To give automakers maximum choice and flexibility, we offer multiple paths
to production. In the first, we offer an affordable solution in both turnkey and
collaborative fusion and policy configurations. In the second, we offer a higher-
performance processor for customers whose particular fusion and policy algorithm
inputs require additional compute power.
Product Brief | Intel and Mobileye Autonomous Driving Solutions

Mobileye EyeQ®5 SoCs

Intel® processors
including proprietary, purpose-built accelerators

EyeQ5 can be delivered in one of two ways: From today’s data-intensive test fleets to tomorrow’s
mass-production vehicles, we provide a range of compute
• Full suite: Complete with hardware-accelerated
density to support diverse paths to market:
algorithms and applications, including sensing, mapping,
fusion, and driving policy software. • The Intel Atom® C3000 processor for automated driving
delivers high performance per watt, packing substantial
• Hybrid/“open” option: Our new software development
compute into low-power designs. It utilizes a power-
kit (SDK) for the EyeQ5 enables OEMs to add their own
efficient microarchitecture on Intel® 14 nm technology. It
functionality for collaborative sensor fusion and driving
is both AEC-Q100 Grade 2 capable and ISO 26262 ASIL C
policy, while taking full advantage of the sophisticated
certified for SEooC (Safety Elements out of Context).
capabilities in EyeQ5. The open option supports four
programming models that enable customers to use the
powerful EyeQ5 SoC to run their own software:
1. OpenCL™
2. Deep learning deployment tools
3. Adaptive AUTOSAR
4. Native Linux* applications

Superior performance
Based on our analysis, compared with NVIDIA’s 1x Xavier*– at roughly the same power. In addition, the Intel and
based solution, a platform featuring an 8-core Intel Atom C3XX4 Mobileye solution meets the highest ASIL D functional safety
SoC and two Mobileye EyeQ5 SoCs is estimated to deliver1: requirements.
• 2.1x higher CPU compute performance5
• 1.6x higher deep learning performance5

8-core Intel Atom® C3XX4 SoC and two Mobileye

EyeQ®5 SoCs vs. NVIDIA 1x Xavier*-based solution1,5

deeP LeArNiNG PerforMANce ANd PoWer totAL cPu coMPute cAPABiLitY

2.4X
Higher DL efficiency
(TOPS/W)
2.1X
Higher
CPU compute
Higher is better Higher is better
2.5 2.5

2.0 2.0

1.5 1.5

1.0 1.0
(Intel Atom®
C3XX4 8C + 2x
0.5 Xavier* EyeQ®5 SoC 0.5 Xavier* EyeQ®5)
0.0 0.0
DL TOPS/W

Figure 1. The Intel-Mobileye solution delivers greater deep learning Figure 2. The Intel-Mobileye solution delivers more than double the
performance efficiency than what was claimed by NVIDIA publicly amount of total CPU compute in the platform compared to NVIDIA’s
for its Xavier* platform.1,2,5 Xavier.1,3,5

2
Product Brief | Intel and Mobileye Autonomous Driving Solutions

8-core Intel Atom® C3XX4 SoC and two Mobileye EyeQ®5

SoCs vs. NVIDIA 1x Xavier*-based solution1,5 (continued)

coMPAriSoN of totAL SoLutioN (eStiMAted)

2.5

2.1X
2.0 Minimum
CPU CPU

1.5
Higher CPU
Minimum
compute
DL DLTOPS

1.0

0.5
Power CPU DL Power 1.6X
Higher DL
performance
1x Xavier* (Intel Atom® C3XX4 8C
estimates + 2x EyeQ®5)

0.0

Figure 3. At the same power, the Intel-Mobileye solution delivers more than twice the amount of CPU compute and 1.6 times higher
performance from a deep learning standpoint.1,4,5

coMPAriSoN of totAL SoLutioN (eStiMAted)

2.5

2.0
Minimum
Power CPU DL CPU CPU

50%
1.5 Minimum
DL DLTOPS

1.0 less power

Power

0.5
2x Xavier* (Intel Atom® C3XX4 8C
estimates + 2x EyeQ®5)

0.0

Performance estimates were obtained prior to implementation of recent software patches and firmware
updates intended to address exploits referred to as "Spectre" and "Meltdown". Implementation of these
updates may make these results inapplicable to your device or system.

Figure 4. Based on estimates, at minimum two NVIDIA Xavier platforms will be needed to meet the level of performance required from a
deep learning and CPU standpoint. The Intel-Mobileye platform can deliver this performance at half the power.1,4,5

3
Product Brief | Intel and Mobileye Autonomous Driving Solutions

Software Built to Sense, Plan, and Act Engineered for Safety with RSS
Our strategy for autonomous driving focuses on delivering Developing established and accepted safety standards
the right tool for each job. This approach is engineered for and policies is critical to making autonomous vehicles a
safety, performance, and affordability. trusted mode of transportation. But how do we define
safety? Showing that we tested the most miles, use the “best
Autonomous driving requires a variety of AI and other tools
practice,” or have the most sensors is simply not enough.
across three stages: sense, plan, and act. It starts with a
Even demonstrating that the autonomous vehicle will not
comprehensive suite of sensors and maps that includes:
break the law doesn’t ensure safety.
• 360-degree surround-view cameras
We must find a way to clearly define—and design for—safety.
• HD maps with REM Global Roadbook™ Only then can we program an autonomous vehicle that will
be smart enough to negotiate the road with unpredictable
• Lidar human drivers—and take a major step toward public
• Radar confidence in autonomous vehicles.

1. Sense That is why we constructed the RSS model. It starts by simply

First, the vehicle must “see” the road. Similar to how formalizing the common sense of human judgment—a
humans drive, we use cameras as “eyes,” the primary central aspect of how humans drive safely. When we think
source of information to analyze the environment in real of what it means to drive safely, we arrive at a few basic
time. Cameras are cost effective, but they also have the principles. For example, always keep a sufficient distance
highest resolution and can see both moving and stationary from the vehicle ahead of you. Or, the right of way is given
objects. Processed camera data is combined with our and not taken. However, those principles are subjective and
high-definition REM Global Roadbook maps to create open for interpretation.
a 360-degree view of the environment with accurate This is where RSS comes in. RSS is a mathematical model
localization to within 10 centimeters. Processed radar of what it means to drive safely. It starts with three main
and lidar are then added for redundancy. definitions:
2. Plan • First, we define a dangerous situation.
Then, the vehicle’s planning module must determine the
• Second, we define what caused the dangerous situation.
next best move. With a simple, semantic-based approach
to decision-making, we eliminate wasted compute. First, • Finally, we define the proper response to that dangerous
a trajectory is planned that is human-like in assertiveness situation.
and feels comfortable to passengers. The RSS safety layer
continuously monitors the situation for any mathematically The RSS model is comprehensive. We analyzed six million
defined “dangerous situations” and initiates a proper accidents in detail. Some 99.4 percent of them fall into one
response if a dangerous situation is caused by another agent. of 37 typical scenarios6, and our mathematical model covers
them all.
By splitting planning into separate modules for comfort
and safety, we achieve the most efficient decision-making Furthermore, by creating a “safety seal” as a separate system
architecture without compromising safety, while allowing within the planning module, RSS frees software developers
opportunities for collaboration and differentiation across our to focus on assertive, comfortable driving dynamics without
OEM customers. needing to account for the safety effects. It also limits safety-
critical errors to the sensing system, which can be empirically
3. Act validated with a reasonable amount of data.
Finally, the decision must be translated into a specific
trajectory for the vehicle to take—in other words, a specific
application of steering, acceleration, or braking. This official
trajectory is calculated by the Intel processor, and then
communicated to various actuators across the vehicle.
4
Product Brief | Intel and Mobileye Autonomous Driving Solutions

iNteL ANd MoBiLeYe AutoNoMouS driViNG SoLutioN ArcHitecture

Fail operational
channel

To actuators
Cameras/
sensors

Trajectory
Vision Fusion/planning validation
and issuance

Radar/lidar

SeNSe PLAN Act

Figure 5. Mobileye EyeQ®5 SoCs are used for computer vision, localization, sensor fusion, trajectory planning, and RSS safety, while Intel®
processors handle translation of the planned trajectory into commands that are issued to the vehicle’s actuation system.

Car-to-Cloud Mapping with REM 1. Harvesting

Camera-equipped L1 and L2 vehicles use EyeQ SoCs to analyze
Autonomous vehicles will need incredibly precise maps the driving path’s geometry and landmarks, compressing
that enable the vehicle to localize itself to a much higher map-relevant data into low-bandwidth (less than 10 KB/km on
accuracy than can be achieved with GPS. These maps are average) Road Segment Data packets that are sent to the cloud.
a requirement to meet functional safety standards for Harvesting begins in 2018 with two million vehicles transmitting
redundancy and must be updated with an ultrahigh refresh data by the end of the year.
rate to reflect changes in the road environment. 2. Aggregation
Current mapping approaches with data collected by In the cloud, Road Segment Data packets are aggregated and
reconciled in a continuous stream to produce a highly accurate
dedicated lidar-equipped survey vehicles are challenging
map, called REM Global Roadbook™.
to scale, due to the sheer number of vehicles that would be
needed to build and update a global map. REM offers a better 3. Localization
way to create maps by harnessing the power of the crowd. In HD maps are sent down to L2+ to L5 vehicles, which use our
the coming years, Mobileye SoCs will be in tens of millions of software to “localize” themselves in the map using low-cost
camera-equipped vehicles worldwide as part of our collision- cameras to detect landmarks. Usage of the Roadbook for
localization begins on several L2+ and L3 vehicle programs
avoidance technology. Mobileye can use these cameras
launching in 2019.
as intelligent agents to generate and rapidly update high-
definition, crowdsourced maps in an affordable way. With REM Global Roadbook, Mobileye will merge data
from multiple automakers, going beyond geofenced areas
With REM, our goal is to create the most accurate maps
and supporting an ultrahigh refresh rate. Using cameras
with the lowest time to reflect reality. REM will also provide
as intelligent data-collection agents also offers new
localization and foresight, critical components to support full
opportunities to monetize other dynamic data for parking,
autonomy. REM works in three steps:
delivery, routing, and infrastructure services.

5
Product Brief | Intel and Mobileye Autonomous Driving Solutions

Trust What’s Inside

Mobileye and Intel are paving the road to autonomous
driving reality with revolutionary software, flexible compute
architectures, purpose-built data center solutions, innovative
approaches to safety, human-machine interfaces that build
trust, and next-generation connectivity that can manage data
needs as drivers become passengers. As the industry’s best
choice for autonomous driving, you can trust what’s inside.
To learn more about the Intel and Mobileye autonomous
driving solution, visit intel.com/automotive.

1. Benchmarks and performance claims:

• Performance tests and ratings are measured using specific computer systems and/or components and reflect the approximate performance of Intel products as measured by those tests.
Any difference in system hardware or software design or configuration may affect actual performance. Buyers should consult other sources of information to evaluate the performance
of systems or components they are considering purchasing. For more information on performance tests and on the performance of Intel products, visit Intel Performance Benchmark
Limitations.
• Results have been simulated and are provided for informational purposes only. Results were derived using simulations run on an architecture simulator. Any difference in system hardware
or software design or configuration may affect actual performance.
2. Deep learning performance and power:
• Deep learning Tera Operations Per Second (DLTOPS) – Typically 1 multiply-accumulate operation = 2 DLTOPS. The widths of the integer matrix multiplication vary by architecture,
dedicated hardware and supported topologies. Any claimed DLTOPS number depends on several assumptions such as frequency, number of MACs and various other hardware
specifications.
• Based on Mobileye EyeQ®5 TOPS performance expectations vs. NVIDIA–claimed Xavier* platform DL performance of 30 TOPS at 30W. Source: http://www.nvidia.com/object/drive-px.html.
3. Total CPU compute capability:
• CPU compute capability defined as the sum of the independently operating systems that are executing SPECint_base2006.
• Based on NVIDIA Xavier* SPECInt_CPU_2000–claimed performance score of 160 from NVIDIA GTC Japan 2016 https://www.gputechconf.jp/assets/files/1062.pdf. SPECInt_CPU_2000
is a retired benchmark. A relevant benchmark is SPECInt_CPU_2006. Estimated SPECInt_CPU_2006 performance scores for Xavier are based on scaling observed from NVIDIA TX2
measurements on SPECInt_CPU_2000 and SPECInt_CPU_2006: similar scaling also observed across similar big core architectures.
• Based on cumulative CPU performance of Intel® solution (Intel Atom® and Mobileye EyeQ®5) platform using estimated SPECInt_CPU_2006 performance scores.
4. Comparison of total solution (sum of all SoCs—not the whole platform):
• Power: Total NVIDIA Xavier* SoC
o DL: NVIDIA–claimed DL power
o CPU: NVIDIA 8C custom CPU power estimate (used a proxy: 8C ARM Cortex A72 power estimate for auto-grade ambient temperature)
o Rest of SoC: Power estimated using 20 percent power to estimate contribution from other components on the chip, based on historical measurements
• Power: Intel
o Intel/Mobileye power estimates based on Intel Atom® C3XX4 8C + 2x Mobileye EyeQ®5
• DL: Based on Mobileye EyeQ5 TOPS performance expectations vs. NVIDIA–claimed Xavier 30 TOPS. Source: http://www.nvidia.com/object/drive-px.html
• CPU: Based on CPU compute capability as defined as the sum of the independently operating systems that are executing SPECint_base2006.
5. Performance estimates were obtained prior to implementation of recent software patches and firmware updates intended to address exploits referred to as “Spectre” and “Meltdown.”
Implementation of these updates may make these results inapplicable to your device or system.
6. “Pre-Crash Scenario Typology for Crash Avoidance Research,” U.S. Department of Transportation National Highway Traffic Safety Administration, April 2007, nhtsa.gov/sites/nhtsa.dot.gov/
files/pre-crash_scenario_typology-final_pdf_version_5-2-07.pdf.
Intel technologies’ features and benefits depend on system configuration and may require enabled hardware, software or service activation. Performance varies depending on system
configuration. No computer system can be absolutely secure. Check with your system manufacturer or retailer to learn more at intel.com.
Tests document performance of components on a particular test, in specific systems. Differences in hardware, software, or configuration will affect actual performance. Consult other sources of
information to evaluate performance as you consider your purchase. For more complete information about performance and benchmark results, visit intel.com/performance.
Intel, the Intel logo, Intel Inside, the Intel Inside logo, and Intel Atom are trademarks of Intel Corporation or its subsidiaries in the U.S. and/or other countries.
*Other names and brands may be claimed as the property of others.
© Intel Corporation
OpenCL and the OpenCL logo are trademarks of Apple Inc. used by permission by Khronos.