As the automotive market evolves toward fully autonomous L5 vehicles, system
designers have faced the incredibly difficult task of harnessing advanced
machine learning (ML) technologies developed in power-hungry, datacenter-class
compute environments and adapting these technologies for embedded automotive
applications with limited power budgets. To achieve this, designers have been
forced to navigate and port complex workflows from one environment to the
other, at considerable expense and delayed time to market, in the absence of a
development platform optimized specifically for AI-based ADAS applications.
NXP endeavored to meet this challenge head on when we announced our highly
anticipated eIQ™ Auto Toolkit in October of last year.
Today, we’re excited to announce that the NXP eIQ™
Auto Toolkit is now available to the automotive mass market. And the need for
this platform couldn’t be greater.
For designers struggling to move from a development environment to AI
application implementations—while simultaneously meeting stringent
automotive safety and reliability standards—the NXP eIQ™
Auto Toolkit gives them the ability to speed up their development cycles
leveraging neural networks, inference engines and NXP’s S32V processor
family to maximize processing agility and performance for real-time AI
These are critical capabilities for a wide range of automated driving apps,
from object classification and path planning to driver/occupant monitoring,
powertrain optimization and more. This AI enablement – made possible in
part by advanced deep learning capabilities for vision, LIDAR and RADAR
technologies – will help to usher in a new era of automotive safety,
intelligence and eco-friendliness.
The Need for Speed
NXP’s eIQ™ Auto Toolkit helps designers convert and
fine tune their AI models leveraging familiar platforms and libraries such as
TensorFlow, Caffe* and/or PyTorch to port their deep learning training
frameworks to high-performance, automotive-grade NXP processors, such as the
S32V processor family. By utilizing pruning and compression techniques, neural
networks can be optimized for maximum efficiency.
One of the key advantages of NXP’s eIQ™ Auto Toolkit
is that it helps designers avoid the cost and time penalties that would
otherwise arise if they had to designate and program onboard compute engines
for each and every layer of a deep learning algorithm—an extremely
cumbersome process. Instead, designers can easily provision task scheduling to
CPU cores and accelerators in a way that matches each individual algorithm
layer to the best suited compute engine. This helps to maximize processing
efficiency while conserving power for other functions.
The resulting boost in performance is enormous. It leads to over 30X higher
performance for given models compared to other embedded deep learning
frameworks, based on NXP’s internal benchmarking using single-thread
TensorFlow Lite models with floating point computation versus an eIQ™
quantized version running on dual APEX-2 vision accelerator cores on an NXP
S32V234. The results speak for themselves and demonstrate how NXP’s
eIQ™ Auto DL Toolkit will help automotive system designers.
With NXP’s eIQ™ Auto Toolkit, these designers also
have the confidence that they’re tapping into a well-established
technology platform. The toolkit is a specially designed evolution of
NXP’s eIQ (“edge intelligence”) machine learning software
development environment. Widely deployed today across a broad range of
advanced AI development applications, NXP’s eIQ software leverages
inference engines, neural network compliers and optimized libraries for easier
system-level application development and machine learning algorithm enablement
on NXP processors.
NXP’s eIQ™ Auto Toolkit is available to designers
today. So if you’ve been frustrated with slow, cumbersome and expensive
development cycles for next-generation AI automotive apps, help is on the way!
(*) Roadmap item
1 Based on Internal NXP benchmarks. Comparisons using single thread Tensor
Flow TF Lite quantized model running on the Arm Cortex-A53 at 1 GHz versus eIQ
Auto version of the model, running on dual APEX2 on S32V234.