AVCC, a global consortium of automotive and technology industry leaders
cooperating on intelligent software-defined and automated vehicle
technology, today announced that its latest technical report is available to
the public for free download: “TR-004 Models and Datasets for Benchmarking
Deep Neural Networks for Automated and Assisted Driving Systems.
“There was no alignment in the industry on how to fairly compare DNN
performance on compute platforms for automated and assisted driving systems.
Two years ago, AVCC set out to create recommendations for a benchmark to
solve this issue,” commented Kasper Ornstein Mecklenburg, Chair of AVCC’s
Micro-Benchmarks Working Group and Staff Performance Analysis Engineer at
Arm. “This is the second of three technical reports on DNN benchmarking and
it addresses the characteristics which make models and datasets suitable for
this purpose, along with a list of recommended models and datasets.”
Having a common view on how to benchmark machine learning (ML) in the
automotive industry benefits the entire ecosystem.
For the automotive and ML communities, this will allow optimized use of
models and datasets that are publicly available, and effectively address the
key usage examples of DNNs in real-world automated and assisted driving
“OEMs and automotive tier 1 suppliers can use the report to have a clear
understanding of test results, allowing them to select the most suitable IP
for their use-case. Plus, IP providers will better understand how to focus
their research resources. It’s a win-win for the industry,” added Paul
Hughes, AVCC Technical Chair and Lead System Architect/Distinguished
Engineer at Arm.
The paper isn’t limited to automotive audiences only: anyone who is involved
in automation, including robotics and other ML applications, can benefit
from the benchmarks outlined in this report.
For those interested in autonomous automotive and assisted driving, the
TR-004 Models and Datasets technical report complements TR-003 Conditions
and Reporting by adding which models, datasets, and automotive context are
relevant for benchmarking.
For more information about AVCC and how it is serving component developers
at automotive OEMs and its suppliers, including how to become a member,
please visit http://www.avcconsortium.org
AVCC is a global autonomous vehicle (AV) consortium dedicated to being the
premier market enabler for intelligent software-defined vehicle technology.
Membership touches upon every facet of the autonomous vehicle and
software-defined vehicle design ecosystem, from technology suppliers to
integrators and beyond. The Consortium serves systems and component
developers at automotive OEMs and its suppliers with strategic programs and
working group publications. AVCC is committed to driving the evolution from
L1 to L5 performance over the next twenty years. At its core, AVCC is
dedicated to providing a vetted architectural design and enabling a
cooperative environment with algorithms and device interfaces for central
and distributed compute for autonomous vehicles.