Edge NVMe Storage for Autonomous Vehicles

Sponsored Post

By Braden Cooper, Product Marketing Manager at One Stop Systems

Adoption of the latest HPC and AI technologies in autonomous vehicles is creating a surge in vehicle real-time data processing capabilities.  The growth in number, speed and resolution of vehicle sensors and related compute performance in autonomous vehicles has in turn led to increased demand for high-capacity, high-throughput storage.  Effective storage within autonomous vehicles must meet three primary criteria: throughput to match the capture and processing rates, a rapid data transfer workflow for offloading the captured data, and a rugged environmental design to operate in any autonomous vehicle conditions.

The compute capacity of an autonomous vehicle system is driven by the power of the CPUs and GPUs therein.  The latest GPUs utilize PCIe Gen 4 with throughput speeds exceeding 20GB/s per GPU.  This raw compute speed has enabled autonomous vehicle system architects to add additional sensors at higher resolutions and maintain a real-time compute capability.  Alongside the capture and compute, the storage elements within the vehicle must maintain the same data rate.  The most rapid form of storage available today is NVMe drives which utilize the same PCIe Gen 4 architecture as the GPUs.  With an optimized PCIe architecture utilizing a mixture of sensor capture, GPUs for compute, and NVMe storage, the entire AI workflow can maintain a balanced, high-performance throughput.

During the development and model training phase of an autonomous vehicle design, the raw data which is captured during travel must be rapidly offloaded from the vehicle to the datacenter where the more powerful model training compute hardware resides. The high-speed, high-resolution sensors could generate hundreds of terabytes per trip.  Cloud-based data transfers of this size made directly from the vehicle add unnecessary delays and costs due to low upload speeds of satellite or 5G networks and costs per byte transferred.  The most effective data transfer of raw data from vehicle to ground station is a physical transfer or a “sneakernet” workflow.  By removing and replacing a full data pack from a vehicle after its travel is complete, the vehicle will be immediately ready for redeployment and the data pack can be securely moved and inserted into a compatible system in the processing facility.

Bringing NVMe storage capacity and speeds to a “sneakernet” workflow in an autonomous vehicle comes with challenges in meeting the environmental conditions of a vehicle.  NVMe is most commonly deployed in rackmount servers integrated in air-conditioned datacenters which experience little to no operational vibration, humidity, dust, or other environmental exposures.  To integrate this form of storage into an autonomous vehicle requires a server design which has been optimized for the structural and thermal challenges of the array of environments the vehicle will experience.  Structural mitigations include using lightweight materials in the chassis design, such as aluminum, and applying structural bracing where needed to reduce the impact of the low frequency resonances seen in various road conditions.  For the thermal design, the system must either integrate with a vehicle’s liquid cooling infrastructure, or provide its own intelligent air-cooled solution based on the conditions the vehicle is exposed to.

As new developments and technologies in HPC are introduced, the need for more capture, compute, and storage in autonomous vehicles will continue to grow.  To meet the throughput required by the latest GPUs, NVMe storage must be integrated into autonomous vehicles in a rugged and easily removable form factor.  One Stop Systems meets the storage requirements of autonomous vehicles by designing rugged optimized servers and PCIe expansion for the AI transportable edge.  The OSS 3U SDS server provides two bulk removable canisters with 8 NVMe drives each for over 240TB of high-speed NVMe storage.  The canisters can be toollessly removed and transported to a ground station for efficient model training.  The SDS’ lightweight and rugged design enable it to operate at the extended environmental conditions which autonomous vehicles experience.  Integrating the latest HPC storage technologies in autonomous vehicles can be a challenge – but a challenge which must be overcome to enable new advances and applications for autonomous vehicles across the industrial and commercial spaces.