Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
Sensor Fusion and Data-in-Motion Processing
for Autonomous Vehicles
Endric Schubert, Ph. D.
CTO
Missing Link Electronics (MLE)
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Disclaimer
2
Presentation Disclaimer: All opinions, judgments, recommendations,
etc. that are presented herein are the opinions of the presenter of the
material and do not necessarily reflect the opinions of the PCI-SIG
®
.
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Acknowledgements
o Ulrich Langenbach, Dir. Eng. – Missing Link
Electronics
o Jim Peek, Dir. Technology – Missing Link
Electronics
3
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Autonomous Vehicles
Better: Automated - not autonomous - driving
4
Click to edit Master title style
5
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019 5
CES 2019: Multi-GPU/SoC ECUs
Click to edit Master title style
6
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019 6
System Overview L4 AV
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Data-in-Motion Processing
o 100 Gbps raw bandwidth, or more
o Data Granularity Issues
7
Pixels Lines Frames
Bytes Kilo Bytes Mega Bytes
Up to 16 Cameras @2.5 Gbps
Up to 4 Radar @ 1..10 Gbps
Up to 4 Lidar @ 1..10 Gbps
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Algos Drive Compute Needs
Combine classical image processing with DNN
8
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
PCIe
®
Comes to the Rescue!
o Relevant CPUs/GPUs/SoCs all have PCIe!
o Build ECU with “PCIe network”
Industry standard, relevant chips all have PCIe
Low latency (micro-seconds)
High bandwidth (tens of Gigabits per second)
9
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
PCIe Non-Transparent Bridge
o Non-Transparent Bridge (NTB) connects multiple Root Ports
o Example of NTB Back-2-Back
(Example from Intel Xeon C5500)
10
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
PCIe NTB – A Defacto Standard
o “Using Non-transparent Bridging in PCI
Express Systems” – Jack Regula, 2004
o Linux NTB from Jon Mason
o Supported by Linux kernel
ntb.h
11
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
NTB - Programmers View
o Great, a network device!
o R/W via TCP sockets
12
Click to edit Master title style
13
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019 13
Automotive Comm. Standards
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
PCIe NTB via Daisy-Chain
o Not optimal for Automotive ECU
Shared Bandwidth
Not resilient to HW failures
Added Latency for ID translation
14
(Example from Intel Xeon C5500)
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
PCIe Non-Transparent Bridge
Network-on-Chip for Any-2-Any Connectivity
between PCIe Roots
15
Click to edit Master title style
16
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019 16
FPGA-Based NTB Architecture
MLE Level-4 AV Platform
Click to edit Master title style
17
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019 17
Data Acquisition & Preprocessing
FPGA technology enables
Flexibility to deal with
sensor i/f
Sensor fusion
Data Acquisition and Data
Preprocessing (DADP)
Data-in-motion
preprocessing
Functional safety (monitor
compute nodes & re-route)
ECU Scale up /
ECU Scale out
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Delivering Performance
Write-Only Communication via Doorbells - NVMe-style
o Avoids difficulties
of multi-device
o Scales to >32 RCs
o Posted Writes
18
Click to edit Master title style
19
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019 19
Peer-2-Peer Communication
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Secure, Reliable, Resilient
o Security / Reliability / Dependability:
single host cannot take entire system down!
local configuration of local properties
receive side memory protection,
incoming traffic for memory space not configured is
discarded and reported (especially useful for (embedded)
devices without IOMMUs)
shielded global NTB config (orthogonal control path for
inter NTB connectivity / routing from Primary NTB to
Secondary NTB)
20
Click to edit Master title style
21
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019 21
Fail-Overs
o ECU-2-ECU via
automotive
10/25/50G Ethernet
o Detect PCIe failure
via AER
o Intra-ECU re-route
o Inter-ECU re-route
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
System-Level Verification
o Use of Xilinx VIP for PCIe + PCIe Root Agents with scripted testcases
o Questa Prime PIPE-Level Sim runs ~ 180 seconds
22
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
System-Level Verification
o Simulation-based performance close to theoretical max:
o
# Run Test Case 60, MLE NTB bandwidth, single peer to peer, TLPs w/ 256 bytes payload, 131072
bytes total
# 213928.00 ns: data transfer, RPA0 to RPA1, duration: 41828.00 ns, bandwidth: 3.133595 GB/s
# 259828.00 ns: data transfer, RPA0 to RPA2, duration: 45896.00 ns, bandwidth: 2.855848 GB/s
# 301644.00 ns: data transfer, RPA1 to RPA0, duration: 41812.00 ns, bandwidth: 3.134794 GB/s
# 347532.00 ns: data transfer, RPA1 to RPA2, duration: 45884.00 ns, bandwidth: 2.856595 GB/s
# 389304.00 ns: data transfer, RPA2 to RPA0, duration: 41768.00 ns, bandwidth: 3.138096 GB/s
# 431092.00 ns: data transfer, RPA2 to RPA1, duration: 41784.00 ns, bandwidth: 3.136895 GB/s
23
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Lab Setup
ProFPGA ZU19 Prototyping System with
multiple x86 as PCIe Root and
cable PCIe Edge-to-Edge, male-to-male, crossed
24
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Results of Performance Analysis
o iPerf – TCP bandwidth measurement (x86)
dummy@buche:~/src/sw/ntbpi$ iperf -s
------------------------------------------------------------
Server listening on TCP port 5001
TCP window size: 85.3 KByte (default)
------------------------------------------------------------
[ 4] local 192.168.1.3 port 5001 connected with 192.168.1.7 port 33686
------------------------------------------------------------
Client connecting to 192.168.1.7, TCP port 5001
TCP window size: 2.01 MByte (default)
dummy@ahorn:~/src/sw/ntbpi$ iperf -c 192.168.1.3 -t 28800 -i 60
------------------------------------------------------------
Client connecting to 192.168.1.3, TCP port 5001
TCP window size: 1.25 MByte (default)
------------------------------------------------------------
[ 3] local 192.168.1.7 port 33690 connected with 192.168.1.3 port 5001
[ ID] Interval Transfer Bandwidth
[ 3] 0.0-60.0 sec 90.9 GBytes 13.0 Gbits/sec
[ 3] 60.0-120.0 sec 91.0 GBytes 13.0 Gbits/sec
[ 3] 120.0-180.0 sec 91.9 GBytes 13.2 Gbits/sec
25
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Conclusion
o To build multi-CPU/GPU/SoC AV ECUs we combine
existing PCIe specification with
defacto industry standards (Linux NTB)
o Single chip solution based on automotive-grade
SoC-FPGA
Delivers performance close to theoretical max.
o Devil in the details:
Interrupt schemes: MSI vs MSI-X
Functional Safety vs Surprise Hotplug
Where to put DMA?
CPU side?
FPGA side? Local write vs remote write?
26
Click to edit Master title style
Copyright © 2019 PCI-SIG
®
- All Rights Reserved
PCI-SIG Developers Conference 2019
Thank you for attending the
PCI-SIG Developers Conference 2019.
For more information please go to
www.pcisig.com
27