NVIDIA DRIVE Thor
Last reviewed
Jun 3, 2026
Sources
14 citations
Review status
Source-backed
Revision
v1 ยท 2,296 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
14 citations
Review status
Source-backed
Revision
v1 ยท 2,296 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA DRIVE Thor (marketed as DRIVE AGX Thor) is a centralized automotive and robotics system-on-a-chip (SoC) developed by Nvidia. It is designed to consolidate functions that previously ran on multiple separate processors, including assisted and autonomous driving, parking, driver and occupant monitoring, the digital instrument cluster, and in-vehicle infotainment (IVI), onto a single computer running NVIDIA's automotive operating system. First unveiled at the GPU Technology Conference (GTC) in September 2022 and later re-cast on NVIDIA's Blackwell GPU architecture, DRIVE Thor is the successor to DRIVE Orin and replaced the cancelled DRIVE Atlan. It serves as the compute element of NVIDIA's DRIVE Hyperion reference platform and is closely related to, but distinct from, the Jetson Thor module for robotics.
NVIDIA introduced DRIVE Thor on September 20, 2022, during the company's GTC keynote, positioning it as a single "centralized car computer" that would unify the cockpit and the autonomous driving stack. In the same announcement, NVIDIA disclosed that DRIVE Thor would replace DRIVE Atlan, a previously announced next-generation SoC that was cancelled before reaching production. Atlan had been planned with a "Grace-Next" CPU and an Ada Lovelace class GPU; rather than ship it, NVIDIA folded its goals into Thor and skipped a product generation.
The lineage of NVIDIA's centralized automotive compute platforms runs as follows:
| Platform | First announced | GPU architecture | Stated AI performance | Status |
|---|---|---|---|---|
| DRIVE Xavier | 2018 | Volta | 30 TOPS | Production |
| DRIVE Orin | 2019 | Ampere | 254 TOPS (INT8) | Production |
| DRIVE Atlan | 2021 | Ada Lovelace (planned) | ~1,000 TOPS (planned) | Cancelled |
| DRIVE Thor | 2022 | Blackwell (revised 2024) | Up to 2,000 FP4 TFLOPS / 1,000 INT8 TOPS | In production rollout |
A notable aspect of Thor's history is that its specifications changed substantially between the 2022 unveiling and the platform that actually shipped. The original 2022 announcement described Thor as combining technology from NVIDIA's Hopper data center architecture (specifically the inference transformer engine), the Ada Lovelace GPU architecture, and the Grace CPU. The Blackwell architecture had not yet been announced at that time. When NVIDIA revealed Blackwell at GTC on March 18, 2024, it re-cast DRIVE Thor as a Blackwell-based platform, describing the chip as built for transformer, large language model, and generative AI workloads. The shipping product is therefore based on Blackwell rather than the Hopper and Ada Lovelace combination cited in 2022.
DRIVE Thor is a unified SoC that pairs an Arm CPU complex with an integrated NVIDIA GPU and dedicated accelerators on a single die, all governed by an automotive-grade safety architecture.
This combination is what NVIDIA calls a single, unified architecture for "advanced driver assistance (ADAS) plus IVI." Rather than dedicating one ECU to the cluster, another to infotainment, and others to the driving and parking stacks, a single Thor SoC can host all of these domains simultaneously, which NVIDIA argues reduces system cost, weight, wiring, and power consumption.
Thor's headline performance figure has been a frequent source of confusion because both the magnitude and the numerical precision format changed over the platform's life. Care is required when comparing the figures.
| Source and date | Stated figure | Precision format |
|---|---|---|
| GTC 2022 unveiling | 2,000 teraflops | FP8 (8-bit floating point) |
| GTC 2022 (alternate framing) | 1,000 teraflops | FP8, introduced as a new automotive data type |
| Blackwell era (2024 to 2026) | Up to 2,000 FP4 TFLOPS | FP4 (4-bit floating point) |
| Blackwell era, equivalent | Up to 1,000 INT8 TOPS | INT8 (8-bit integer) |
The current official specification for the Blackwell-based DRIVE AGX Thor SoC is up to 2,000 FP4 TFLOPS of compute, which NVIDIA equates to 1,000 INT8 TOPS for deep learning inference. The shift in the headline number from FP8 to FP4 reflects Blackwell's support for 4-bit floating point precision; a given chip produces a larger teraflops number at FP4 than at FP8 because each operation handles narrower data. As a practical comparison, NVIDIA cites the 1,000 INT8 TOPS figure as roughly four times the 254 INT8 TOPS delivered by DRIVE Orin. Published per-vehicle figures for early production cars have varied between sources, with some citing approximately 700 TOPS and others 1,000 TOPS for a single-Thor configuration, so individual vehicle numbers should be read against the specific quoted source rather than treated as a single fixed value.
DRIVE Thor runs NVIDIA DriveOS, the company's safety-certified operating system for autonomous vehicles; the version paired with Thor developer kits is DriveOS 7 (with releases such as DriveOS 7.0.3). DriveOS provides secure boot, a safety-oriented hypervisor and real-time operating system, CUDA and TensorRT for accelerated inference, and the NVIDIA DriveWorks SDK with libraries for perception, sensor fusion, and vehicle-systems integration. TensorRT 10 on Thor adds support for NVFP4 precision and INT4 weight-only quantization tuned for the Blackwell GPU.
On top of DriveOS, NVIDIA offers application-layer software, historically branded DRIVE AV (the autonomous driving stack covering perception, mapping, planning, and control) and DRIVE IX (the intelligent-experience and in-cabin software for driver and occupant monitoring and infotainment). The platform is built to run modern transformer-based perception models, vision-language-action (VLA) models, and generative AI in the vehicle, which is the central rationale for moving to the Blackwell architecture and its transformer engine.
NVIDIA ships two product lines under the "Thor" name that share the same underlying Blackwell-based SoC family but target different markets and software stacks:
| DRIVE AGX Thor | Jetson Thor | |
|---|---|---|
| Primary market | Automotive and autonomous vehicles | Robotics and physical AI |
| Software stack | NVIDIA DriveOS, DRIVE AV / DRIVE IX | NVIDIA Isaac, Holoscan, Metropolis |
| Safety focus | ISO 26262 ASIL-D, ISO 21434 automotive | Robotics and edge deployment |
| Form factor | In-vehicle compute platform and dev kit | T5000 system-on-module and AGX dev kit |
Both are built on the Blackwell GPU architecture and the 14-core Arm Neoverse V3AE CPU. DRIVE Thor is purpose-built for vehicles and carries automotive functional-safety certification readiness, while Jetson Thor is aimed at humanoid robots, autonomous mobile robots, and industrial systems and runs NVIDIA's robotics software. NVIDIA brought the Jetson AGX Thor developer kit and the Jetson T5000 production module to general availability on August 25, 2025, with the developer kit priced from $3,499 and the T5000 module from $2,999; that module is rated at up to 2,070 FP4 TFLOPS with 128 GB of LPDDR5X memory. The two lines are distinct products and should not be conflated, even though their compute cores are common.
NVIDIA has announced a broad roster of automakers and autonomous-driving developers committed to DRIVE Thor. Many of the earliest and most prominent design wins are Chinese electric-vehicle and automated-driving companies, alongside a growing list of global manufacturers and trucking firms.
Announced passenger-vehicle and EV adopters include:
Announced autonomous-driving and trucking partners include Nuro (Level 4 passenger autonomy), Plus and Waabi (autonomous trucking), and WeRide (Level 4 commercial robotaxi and logistics applications).
The Geely-affiliated Lynk & Co 900 is widely reported as the first mass-production vehicle to ship with a DRIVE Thor SoC. It was unveiled on January 3, 2025, with sales in China beginning on April 28, 2025; in that vehicle the higher trims use DRIVE Thor while lower trims use the prior-generation DRIVE Orin-X.
At CES in January 2025, NVIDIA, autonomous-trucking company Aurora, and tier-one supplier Continental announced a long-term partnership to build driverless trucks at scale on DRIVE Thor, with Continental integrating Thor and DriveOS into the Aurora Driver hardware. The companies indicated that DRIVE Thor production samples would arrive in the first half of 2025, with high-volume hardware manufacturing targeted for 2027; the Volvo VNL Autonomous truck was shown as part of the effort.
DRIVE Thor is the compute core of DRIVE Hyperion, NVIDIA's production-grade autonomous-vehicle reference architecture. The DRIVE AGX Hyperion 10 generation pairs two in-vehicle DRIVE AGX Thor platforms (each delivering 2,000 FP4 TFLOPS, equivalent to 1,000 INT8 TOPS) with the safety-certified DriveOS and a qualified multimodal sensor suite reported as 14 high-definition cameras, nine radars, one lidar, and twelve ultrasonic sensors.
On October 28, 2025, NVIDIA announced a partnership with Uber to scale a global fleet of Level 4 autonomous, Hyperion-ready vehicles, with Uber targeting on the order of 100,000 vehicles as it begins scaling its autonomous fleet from 2027. NVIDIA named a set of Level 4 development partners building on or compatible with Hyperion 10: Stellantis, Lucid, and Mercedes-Benz for passenger mobility; Aurora, Volvo Autonomous Solutions, and Waabi for long-haul freight; and an ecosystem including Avride, May Mobility, Momenta, Nuro, Pony.ai, Wayve, and WeRide.
When first announced in 2022, DRIVE Thor was targeted at automakers' 2025 model-year vehicles, and the 2024 Blackwell-era messaging reiterated production "as early as next year" (2025). In practice the timeline played out across 2025 and beyond: the DRIVE AGX Thor developer kit opened for pre-order with deliveries beginning in September 2025 (offered in benchtop and in-vehicle configurations), production samples for the Aurora and Continental trucking program were slated for the first half of 2025, and the first production passenger vehicle, the Lynk & Co 900, reached the Chinese market in the second quarter of 2025. Large-scale deployments tied to robotaxi and autonomous-trucking programs are oriented toward 2027 and later, reflecting the multi-year qualification cycles typical of automotive-grade hardware.
DRIVE Thor represents NVIDIA's bid to make a single, software-defined computer the central nervous system of next-generation vehicles. By unifying the driving stack, parking, monitoring, cluster, and infotainment on one Blackwell-based SoC, it aims to replace a patchwork of domain controllers with one high-performance, functionally safe processor, lowering bill-of-materials cost while providing enough headroom to run large transformer and generative AI models on board. Its broad adoption among Chinese EV leaders, global automakers, and autonomous-trucking and robotaxi developers, together with its role as the compute foundation of the DRIVE Hyperion platform and its kinship with the Jetson Thor robotics module, positions it as a central piece of NVIDIA's strategy for physical AI in both vehicles and robots. The platform also illustrates how rapidly NVIDIA's automotive roadmap tracks its data center architectures, given that Thor was re-based from Hopper and Ada Lovelace onto Blackwell between its announcement and its production launch.