Why do the world’s leading autonomous driving companies all adopt the VAST Data platform?
Release time:
2026-06-02 10:53
Source:
In Autonomous driving Companies that stand out in industry competition are not merely refining their algorithms; they are also reimagining the data platforms that underpin connected and autonomous vehicles and their broader ecosystems.
Enterprises that stand out in industry competition, It’s not just about optimizing algorithms; it’s also about rebuilding the data platform that underpins connected and autonomous vehicles and their ecosystem. 。
The challenge of data scale continues to intensify.
A test vehicle equipped with multiple sensors can generate up to several terabytes of data per day from sources such as cameras, LiDAR, millimeter-wave radar, and ultrasonic sensors. If this volume is scaled across a global fleet of data‑collection vehicles, combined with synthetic data generation and simulation workloads, the total data management requirements could reach hundreds of petabytes.
However, The challenges of ADAS data infrastructure lie not only in the sheer volume of data, but also in the highly diversified access patterns across the entire business workflow. 。
The scenario of collecting raw sensor data is entirely different from the scenario of randomly streaming video clips to a GPU cluster for model training. Similarly, the performance characteristics required to run validation workloads in safety-critical edge use cases differ markedly from those needed to synchronize optimized model outputs back to vehicles via over-the-air (OTA) updates.
Moreover, Today, the platform must also support real-time remote data processing for intelligent connected vehicles on the road. , these inexhaustible Telemetry data It will be used both for analytics dashboards and as a data source for training next-generation models.
Each stage in this business workflow has distinct requirements for throughput, latency, and data compliance governance. Traditional architectures often compel teams to build and maintain separate systems for each use case, resulting in severe data silos and fragmentation across global R&D centers. When data is scattered across incompatible, standalone systems, the sheer volume of data makes integration virtually impractical. 。
Data is not only massive in scale but also excessively fragmented, making it difficult to flow efficiently across sites and between on-premises environments and the cloud. Moreover, data fragmentation prevents the establishment of a unified… “Single Source of Truth” , which has also slowed down the efficiency of collaboration between global R&D personnel and data scientists.
The end result is that enterprises are left cobbling together a fragile architecture: high-performance computing file systems, self-managed object storage, and cloud‑based caching environments, barely kept afloat by a patchwork of custom ETL (extract, transform, load) workflows. The R&D team has devoted substantial engineering effort to data‑flow governance, leaving them with little time to optimize the models themselves. 。
Traditional infrastructure struggles to support the massive scale of ADAS operations.
The industry as a whole faces the same predicament: enterprises that initially adopted traditional HPC file systems such as GPFS were able to meet their needs in the early stages of their operations, but… As data volumes grow to hundreds of petabytes, operational complexity surges and architectural stability deteriorates. 。
These “no‑shared” parallel file systems are designed specifically for the sequential I/O patterns required by traditional HPC simulations, rather than for the random I/O demands of deep learning. As a result, this leads to severe… I/O starvation , that is Expensive GPU clusters sit idle because the storage layer simply cannot deliver data fast enough to keep them running at full capacity. 。
The hastily added, standalone S3‑compatible object storage introduced to address these issues has in turn created new protocol silos. Faced with the need to scale metadata for billions of small files—such as single‑frame annotated video clips—the only option is to overprovision storage capacity to avoid performance bottlenecks.
A global autonomous‑driving software company encountered a similar situation. At present, the company has deployed hundreds of NVIDIA H100‑class GPUs and over 100,000 CPU cores across on‑premises data centers and cloud environments; however, its existing GPFS system combined with a self‑built S3 cluster is unable to meet the demand— As capacity scales, operational complexity rises in tandem; each business expansion entails managing more systems, synchronizing additional data replicas, and troubleshooting a growing number of potential failure points. 。
As a result, the storage architecture has failed to keep pace with the company’s investment in computing power, severely impacting GPU utilization.
Another ADAS software vendor specializing in the security domain faces a different yet related challenge. Its team is undergoing a technological leap, transitioning from training perception models on static images to training them on dynamic video. This shift has led to a corresponding surge in both system throughput and data volume.
Since hundreds of GPUs need to be shared among a large research team comprising hundreds of data scientists, Their existing infrastructure cannot scale its performance independently to keep pace with business growth. . Metadata performance bottlenecks frequently cause service disruptions. Moreover, the legacy compression algorithms they continue to use achieve only a 1.1:1 data reduction ratio on JPEG and PNG training datasets, meaning their storage costs grow almost linearly with data volume.
The true core requirements of the ADAS business workflow
Setting aside the self‑promotional claims of various vendors, a closer look at the genuine needs of autonomous‑driving R&D teams reveals that their requirements are remarkably aligned:
-
Stable, predictable throughput , The speed must reach the level of tens of GB/s. , to ensure that the GPU can operate at full capacity during training, thereby completely addressing the I/O‑bound bottleneck commonly seen in traditional architectures;
-
Native streaming data collection capabilities , it natively handles daily terabyte-scale sensor and connected‑vehicle data, is compatible with the Kafka interface, and enables a single platform to both store training data and serve as an event bus for real-time vehicle telemetry;
-
Supports multi-protocol access with NFS and high-performance S3. , one dataset simultaneously supports both HPC and cloud-native workloads, eliminating the need for data replication and migration;
-
Performance, capacity, and metadata can be scaled horizontally and independently. , avoid over-allocating other resources in order to enhance a particular capability;
-
It remains effective even for compressed formats such as JPEG, PNG, and encoded video. Efficient data reduction capability ;
-
The global namespace can override both the local cluster and the public cloud environment. , to support sudden spikes in computing demand. This approach—“letting computation find the data rather than having the data seek out computation”—addresses the issue of data fragmentation;
-
Supports cross-site data sharing , eliminating the need to replicate entire petabyte-scale datasets across sites.
No conventional system can simultaneously meet all of the above requirements. That’s why leading ADAS companies are consolidating their operations onto a unified data platform. Moreover, this platform is built from the ground up specifically for AI workloads.
Real-World Implementations by Leading Autonomous Driving Companies
The aforementioned global autonomous driving software company has integrated its existing GPFS with… Self-built S3 cluster The entire infrastructure has been migrated to VAST Data AI OS, integrated with NVIDIA DGX SuperPOD, to support ADAS and large‑model R&D, delivering immediate, tangible results:
-
A unified global namespace that spans NFS and high-performance S3. , eliminating protocol silos and redundant data storage;
-
VAST Data’s disaggregated all‑shared architecture (DASE) enables all CPUs and GPUs to simultaneously access all SSD data, Completely eliminates the east-west traffic bottleneck of traditional shared‑nothing architectures. , enabling true parallel data access;
-
Code repository Data reduction ratio as high as 12:1 ;
-
Elastic scaling capability in the cloud platform , within a single namespace, compute workloads can be dynamically scheduled to the public cloud, with the computation results synchronized back to the on-premises environment.
Today, the company’s effective capacity on the VAST Data platform has exceeded 10 PB, with plans to double its scale within two years. It has also developed a real-time mapping system that enables centimeter-level precision in identifying and sharing road‑condition risks among vehicles. Today, we have infrastructure capable of keeping pace with the rapid growth of massive data volumes. 。
Another ADAS software vendor specializing in safety has also reaped substantial benefits:
-
VAST Data for its training operations Stably delivers throughput of up to 50 GB/s. , ensuring that its GPU cluster operates at full capacity and eliminating the “starvation” caused by insufficient data supply;
-
The platform’s advanced similarity-based data reduction algorithm, Achieved a 2.5:1 compression ratio on JPEG and PNG datasets. , compared with the traditional 1.1:1 reduction ratio, it more than doubles the effective capacity;
-
The DASE architecture supports compute nodes and storage nodes. On-demand independent scaling and seamless online upgrades, with compute capacity expansion unrestricted by storage capacity limits. , thereby completely resolving the original metadata performance bottleneck and thoroughly addressing the frequent business interruptions they experienced on their legacy infrastructure.
After deploying VAST Data clusters across multiple data centers, enterprises can leverage the VAST DataSpace global snapshot feature to… It enables seamless sharing of complete datasets across locations without the need for data replication, fundamentally addressing data silos and enhancing cross‑regional collaboration efficiency. 。
R&D personnel can access the data they need at any site, quickly generate datasets for training model variants, and collaborate seamlessly across regions—without worrying about data redundancy or bloat. For a large team of hundreds of data scientists who must compete for GPU resources in a shared scheduler, This frictionless data-access capability has become a core competitive advantage. 。
From model training to real-time inference deployment
Discussions about ADAS often center on model training, but training is only one part of the process. The models that underpin scene understanding, behavior prediction, and path planning must ultimately be deployed in production, continuously processing streaming sensor data and making real-time decisions.
Therefore, Training infrastructure must simultaneously support inference workloads. Moreover, data exchanged among vehicles, edge nodes, and data centers must be collected, processed, and scheduled in real time as a continuous stream, rather than through the traditional overnight batch-processing approach.
This is precisely the core advantage that enables VAST Data’s architecture to outperform traditional storage. : VAST Database can serve as Real-time event streaming platform It is compatible with Kafka and natively supports the Apache Arrow interface, enabling large-scale ingestion and querying of connected‑vehicle and sensor streaming data. The same platform that handles petabyte‑scale training data can simultaneously ingest real‑time vehicle telemetry, perform real‑time operational analytics, and support inference workloads—eliminating the need to build a separate streaming architecture.
For ADAS R&D teams, this means that the closed-loop process—from data collection and model training to validation—no longer stops at model export. Perception models for identifying construction zones, decision‑making models for planning detour routes, and prediction models for anticipating the behavior of surrounding vehicles, All rely on the same underlying data platform to rapidly retrieve contextual data. 。
In addition, OTA update verification, digital twin simulation, and real-time fleet‑operation dashboards all eliminate the need to build new architectures from scratch. It can run directly on the same VAST Data AI OS that supports model training. 。
This end-to-end, full‑stack capability is precisely what sets a data platform—akin to an operating system—apart from conventional storage systems. ADAS R&D cannot be confined to training models in simulation environments; it also requires a unified infrastructure to support the model’s entire lifecycle—from validation and deployment to continuous iteration in real-world scenarios—all of which must be built on a single, integrated platform.
Industry Macro Trends
ADAS is merely a microcosm of a larger technological transformation. Beyond the automotive industry, robotics firms, industrial automation companies, and drone‑swarm operators all face the same core challenge: Building AI systems that can perceive, reason, and act in the physical world. Its data volume and complexity far exceed the upper limits of traditional infrastructure design.
Current architectural requirements for ADAS — Capabilities such as massive streaming data ingestion, real-time inference, global data access, and continuous model iteration. , is becoming a standard foundational capability for all companies building “physical AI.”
Only organizations that view data infrastructure as a strategic investment—rather than merely a cost center—can deploy safer, smarter systems more quickly. The autonomous driving industry has demonstrated that the AI operating system underlying a model is just as critical as the model itself. 。
Author of this article: Aaron Chaisson, Vice President of Product and Solutions Marketing at VAST Data.
Recommended News
GTC25 | The Chinese-subtitled version of NVIDIA CEO Jensen Huang’s keynote is now available!
Watch NVIDIA CEO Jensen Huang unveil the latest developments in AI and accelerated computing at GTC 2025.
Tsinghua Unigroup Digital joins forces with VAST Data to usher in a new era of AI storage.
Ziguang Xiaotong Technology Co., Ltd. has officially become the exclusive distributor of VAST Data in China.
Why do the world’s leading autonomous driving companies all adopt the VAST Data platform?
In the fiercely competitive autonomous‑driving industry, companies that stand out are not merely refining their algorithms—they are reimagining the data platforms that underpin connected autonomous vehicles and their broader ecosystem.
In October 2022, Ziguang Xiaotong, a subsidiary of Ziguang Digital Group, signed an agency agreement with NVIDIA. Ziguang Digital officially became the general agent of NVIDIA Data Center GPU,NVIDIA DGX System, NVIDIA Professional Visualization, NVIDIA Virtual GPU and NVIDIA Network Products in China. Start a new chapter of cooperation with NVIDIA in the field of artificial intelligence.
ChatGPT is an AI dialogue model developed by OpenAI in the United States in November 2022. Due to its excellent text generation and dialogue interaction capabilities, it has rapidly become popular worldwide, with more than one million active users in 5 days and more than 100 million active users in 2 months.