Inspur began work on autonomous driving around late 2021, initially focusing on the algorithm side. In October 2022 Inspur proposed the vision-only DABNet4D, and in March 2023 it proposed IEI-BEVFusion++. On the nuScenes detection leaderboard the system achieved a high NDS of 77.6%, one of the leading scores in 3D object detection.
Background
Automotive high-performance compute can be considered edge computing. Inspur is a leading edge computing server vendor in the Chinese market, and the company's server expertise naturally extends into the automotive autonomous driving domain.
In August 2023 Inspur released the EIS400 autonomous driving domain controller together with the AutoDRRT compute framework. The product differs significantly from many existing automotive domain controllers.
EIS400 domain controller
System architecture
Inspur uses four NVIDIA Orin modules for AI acceleration and coprocessing. These four AI accelerators and coprocessors each run distinct tasks: each task domain is handled by a single Orin. In other words, this is not four Orins running in parallel to increase raw throughput; instead they act as largely independent coprocessors arranged in a near-serial task partitioning. This differs from many traditional automotive domain controllers.
As a server vendor, Inspur uses a PCIe switch to connect the four Orin modules. In contrast, many automotive vendors prefer Ethernet switches for cost reasons. Mainstream Gigabit Ethernet provides only 0.125 GB/s per link, while NVIDIA NVLink used between GPUs can reach about 900 GB/s. Given that gap, paralleling four Orins to obtain a fourfold compute increase is effectively impossible; at best a modest gain is expected. Inspur selected a third-generation PCIe switch offering a cost-effective solution: the Orin connections use 8 lanes with 8 GB/s bandwidth, and the CPU connection uses 16 lanes at 16 GB/s. PCIe has evolved to Gen6, where single-lane bandwidth can reach 64 GB/s, roughly eight times Gen3, though at significantly higher cost. Inspur also includes an Ethernet switch with eight Gigabit Ethernet ports and three SGMII interfaces to connect up to four lidars. While 100 Mbps Ethernet can be marginally sufficient, Inspur chose Gigabit ports for forward compatibility.
System management and security
BMC system monitoring refers to the baseboard management controller commonly used in servers. The BMC is a dedicated chip or embedded processor with its own network interface for direct network connection. Through the BMC, system administrators can remotely monitor, maintain, update, and control the system. BMC provides system management firmware, utilities, peripherals, and sensors, and supports features such as KVM (keyboard, video, mouse), power management, virtual media, SNMP, logging, and remote access. Reliability expectations in the server domain are comparable to those in automotive applications.
For security MCU functionality Inspur uses the Infineon TC397. Inspur also added a CPLD to handle positioning signals, referred to as PPS here (Precise Positioning Service). This is not the GPS PPS pulse-per-second. A CPLD can be thought of as a compact FPGA.
Inspur brought server-grade technology into automotive autonomous driving. The server model of multi-accelerator plus CPU is retained: four NVIDIA Orin modules are treated analogous to four discrete GPUs, while the CPU remains an x86-based system.
ROS, Autoware and middleware
ROS is the most common operating framework in the robotaxi field; apart from a few exceptions, many systems use ROS. ROS originated from Stanford AI Lab and the robotics company Willow Garage. After 2008 the Open Source Robotics Foundation (now Open Robotics) took a major role in maintenance. ROS2 was released at the end of 2016 and continues to be updated roughly annually; the latest stable release in 2023 was Iron Irwini. A well-known open-source autonomous driving stack built on ROS2 is Autoware. At its core, ROS provides a communication middleware for sensor and node interaction.
Autoware was first released in August 2015 by a research group at Nagoya University led by Prof. Shinpei Kato. Prof. Kato later founded Tier IV to maintain and apply Autoware for real autonomous vehicles. Today Autoware is available in major branches: AutoWare.ai (ROS1) and AutoWare.auto / Autoware.universe (ROS2). The framework supports modules for localization, mapping, object detection and tracking, traffic light recognition, mission and motion planning, trajectory generation, lane detection and selection, vehicle control, sensor fusion across cameras, lidar, and radar, deep learning components, rule-based systems, connected navigation, logging, and virtual simulation tools.
The Autoware Foundation maintains Autoware. Its members include major semiconductor and automotive companies.
In the EIS400 partitioning, the four Orin modules and the CPU have distinct responsibilities. One SoC is dedicated to lidar processing, specifically Velodyne lidar in the presented configuration. The CPU mainly manages the Autoware base stack.
After optimizations the latency was significantly reduced.
Inspur did not adopt Autoware unchanged; it introduced DDS as part of the middleware stack. DDS is a communication middleware targeted at high-reliability environments. ROS network communication supports TCP and UDP; TCP is the default transport.
TCP in brief: after the sender transmits a message, the receiver must send an acknowledgment to the sender, confirming receipt. If the sender does not receive the acknowledgment, subsequent messages may be withheld.
UDP in brief: the sender transmits messages without requiring acknowledgments from the receiver. UDP is often suitable for autonomous driving telemetry, though commercial-grade autonomous systems generally require more sophisticated transport features.
Communication middleware commonly supports point-to-point, message-queue, and publish/subscribe patterns. SOME/IP and DDS adopt the publish/subscribe model. Compared with signal-oriented CAN, SOME/IP and DDS are service-oriented protocols. SOME/IP (Scalable service-Oriented MiddlewarE over IP) is a service-based transport standard originally developed by BMW and integrated into AUTOSAR. Commercial SOME/IP implementations are supplied by vendors such as Vector, while an open-source SOME/IP is maintained by GENIVI. Traditional automakers commonly use SOME/IP.
DDS stands for Data Distribution Service and is a distributed communication standard from OMG that uses a publish/subscribe model and supports multiple QoS policies.
DDS models distributed data as "topics" and defines data producers and consumers as "publishers" and "subscribers." Nodes are logically peer-to-peer, and connections can be point-to-point, point-to-many, or many-to-many. With QoS controls, DDS enables automatic discovery and configuration of network parameters.
DDS was first applied for complex upgrade compatibility in U.S. Navy ship networks and later expanded into aerospace, defense, communications, and automotive fields. In 2018 DDS was introduced into AUTOSAR AP as an optional communication method. ROS2 and Cyber RT use open-source DDS implementations as their core transport. NVIDIA SoCs such as Xavier and Orin provide DDS interfaces.
Globally, RTI (Real-Time Innovations) is the largest DDS vendor, with a market share around 80%. RTI led the DDS standard development and provides a commercial DDS product named Connext DDS. A notable open-source DDS implementation is eProsima Fast DDS, developed by a team with RTI roots. eProsima released Fast DDS source code on GitHub, though paid support is available. Due to cost considerations, NVIDIA drivers often expose Fast DDS or OpenDDS in their Driver.OS. RTI views open-source DDS as its primary competitor.
AutoDRRT fault tolerance
In November 2023 Inspur participated in the Japan Automotive AI Challenge and won first place using EIS400 and AutoDRRT. The Japan Automotive AI Challenge is an international autonomous driving competition that challenges teams to solve a range of industrial autonomous transport scenarios. The 2023 event focused on factory autonomous transport and included obstacles, smoke interference, S-shaped and narrow L-shaped paths. The team that drove the farthest distance won. The competition attracted over 50 top autonomous driving teams, including university and industry teams. After online simulation qualifiers, 17 teams advanced to the onsite finals, including both academic and corporate teams.
