Automotive test system design for heterogeneous sensor fusion
01 April 2020
Today, hardware solutions for vehicle testing or the development of driver assistance systems & automated driving are required to perform increasingly complex data processing & data logging functions. An apt example is the various sensors for situational awareness & AI inference systems that need integrating, and often require highly intensive GPGU processing.
This case study was originally featured in the April 2020 issue of EPDT magazine [read the digital issue]. Sign up to receive your own copy each month.
To address this need, as Korbinian Hecker, Automotive Product Manager at InoNet tells us, the developer & manufacturer of industrial computer systems has now extended its wide range of hardware solutions for automotive test systems with the first AMD Ryzen Embedded V1000 processor based platform.
The development of advanced driver-assistance systems (ADAS) and self-driving cars – for example, in compliance with ISO 26262 – requires continuous testing and validation. And with the increasing complexity of equipment, the proportion of hardware and software in the vehicle is growing. As a result, the test effort en route to series maturity also increases. Wherever possible, the required often highly complex computing processes should be outsourced to HIL, SIL and VIL (hardware-, software- and vehicle-in-the-loop) to ensure rapid, cost efficient and reproducible validation. For this purpose, they must provide flexibly interchangeable performance and interface scalability – which the ecosystems of the x86 processor technology are particularly well suited for.
Exponentially increasing data volumes
Between driving automation levels 2 and 5, the amount of data increases exponentially. High write rates and time synchronisation of data – partly also of raw data – are required to efficiently account for big data application during development. Besides familiar environmental sensors such as ultrasound and radar, more and more video cameras are now being used for 360° surveillance, in addition to laser and LIDAR systems for autonomous driving, whose data must also be recorded and evaluated.
And of course, the outputs to actuators such as steering, engine and transmission control, as well as light and warning sensors, must also be integrated. The Audi A8’s traffic jam pilot, which uses 24 different sensor systems, including five radar sensors and six cameras, illustrates how high the sensor density for Level 3 systems is, even today. In the future, vehicle-to-X communication via mobile radio networks, such as the new 5G network, will need to be added. It will be used to communicate with the systems of nearby road users and stationary roadway facilities, such as traffic monitoring systems, traffic lights and parking premises.
Besides all this, the vehicle hardware is exposed to high temperatures, severe shocks and vibrations during the tests and must be able to withstand these stresses, not only throughout reliable continuous operation, but also during test drives in sometimes extreme environments.
Heterogeneous requirements for automotive test systems
One company that specialises in the design of such automotive test systems is InoNet, a vendor of customised industrial computing solutions. Offering enormous computing power and industrial-grade ruggedness for these tasks, the embedded systems from InoNet are optimally designed for use in vehicles, and can easily withstand high temperatures, shocks and vibrations. Scalable data volumes of up to 256 TB and data rates of up to 12GB/s make the in-vehicle PCs ideal for rapid control prototyping applications. Easily replaceable hard disks, a wide-range power supply (with ignition control) and time synchronisation in case of cascading (hardware time stamping) provide a particularly comfortable and efficient user experience.
However, up until now, automotive test experts focusing on sensor fusion have been lacking one thing to connect and process pretty much everything required today, from the simple LIN and evolving CAN-FD bus to GigE vision or USB 3.0 cameras: an embedded computing platform for extremely powerful heterogeneous parallel processing based on freely programmable application processors. For this purpose, InoNet has now developed a system that isn’t – for the first time – based on Intel, but on AMD embedded processor technology.
Heterogeneous parallel processing with GPGPUs
The most important advantage of this new design is the possibility to offer ultra-efficient parallel processing for data preprocessing of situational awareness sensors and AI inference algorithms on the basis of today’s most powerful processor integrated graphics units (GPU). These graphics units are also known as general purpose graphics processing units (GPGPU), since they don’t just process graphics data for screen display extremely efficiently, but also all other types of data. Compared to CPUs, they offer many more computing units for parallel data processing.
The InoNet system, for example, features GPGPUs with up to 704 compute units that can process 704 data points in one go. By comparison, a quad-core processor with simultaneous multiprocessing can only process up to 8 data points in parallel. A further advantage is the fact that there are discrete GPUs of the same generation as the processor with integrated graphics, which offers unified management and high scalability. Another convincing factor is that the AMD GPU ecosystem extends all the way to server farms that have been developed precisely for the most powerful deep learning systems. They therefore also offer a perfect basis for more massive, GPGPU-based parallel processing with highest scalability, based on a single ecosystem that also supports open source standards such as ROCm and Tensorflow.
In recent years, AMD has overall achieved an outstanding position both in the server segment and in embedded systems. This is due to the Zen microarchitecture and Infinity bus, which allows AMD to integrate the same core technology into all of its Ryzen and Epyc processor families. Developers and users benefit from extremely high compatibility for efficient software re-use, seamlessly ranging from data centres to embedded edge servers and deeply embedded systems.
Automotive test systems featuring the AMD Ryzen Embedded V1000 series
For its first automotive test platform with AMD embedded processor technology, InoNet chose the AMD Ryzen Embedded V1000 series, which offers a highly scalable TDP from 12 to 54 W. As an accelerated processing unit with central CPU and GPU integration, it gives automotive developers, above all, high performance at an attractive price. In detail, the AMD Ryzen Embedded V1000 processor offers (according to benchmark testing done at AMD’s Embedded Software Engineering Lab) up to 2x more performance than its predecessor and up to 46% more multi-thread performance than competitive solutions. The new AMD Ryzen Embedded processors also provide significant gains in graphics – always a strength of AMD processor technology. They now offer more than twice as much graphics performance than the previous accelerated processing unit (APU) of the AMD Embedded R series (codenamed Merlin Falcon) and even up to three times more graphics performance than directly comparable competitive solutions (based on the 3dMark® 11P benchmark). All in all, the new AMD Ryzen Embedded V1000 APUs with Zen CPU and Vega GPU achieve a performance throughput of up to 3.6 TFLOPS.
Flexible system design for the fusion of heterogeneous sensors
InoNet uses the iBASE Mini-ITX MI988 motherboard to implement this processor generation into its automotive test systems, which can be customised for a wide variety of vehicle test and development requirements. Soldered processors and secure external interfaces such as M12 connectors and D-sub sockets with up to 25 pins guarantee high ruggedness and secure connectivity. Their signal lines can be equipped with CAN and CAN-FD, LIN, RS232 and any other vehicle networks for communicating with typical vehicle bus systems.
When used as a data logger, up to 8TB SSDs can be output quickly and easily via up to 2 shuttles. For inter-system communication, the industrial PC provides 4x 10 GbE interfaces, via RJ45 connectors for cost-efficient copper-based connectivity. The industrial box PC offers two PCIe x8 slots for individual extensions, for example with dedicated GPGPUs. The configurable automotive power supply with lockable Neutrik plug with ignition pin provides 250 W of continuous power, making it possible to supply even power-hungry high-performance GPGPU cards. The intelligent cooling solution ensures safe HPC operation even at ambient temperatures of 10 to 60°C.
Overall, customers benefit from a solution that enables flexible sensor data fusion in a single system.
A potential candidate for innovative target systems
With increasing autonomy comes a need for ever more powerful processors, even in OEM series. There are various model studies presenting 10 GbE and more with processors at embedded server level. Potentially, AMD embedded processor technology could therefore even be integrated as a runtime processor, in a specifically customised design, into a series’ target system. In the past, AMD has demonstrated the ability to customise its platform at the silicon level for game consoles, enabling it to develop and manufacture OEM solutions in very specific designs, something that is particularly important for the automotive sector.
Contact Details and Archive...