NVIDIA DRIVE Thor (1000 TOPS Autonomous Driving)
Overview
NVIDIA DRIVE Thor is NVIDIA's 4th-generation automotive-grade autonomous driving SoC, announced at GTC 2022-09, taped out in 2024, Developer Kit GA August-September 2025, ZEEKR production starting 2025 H2. 4nm Samsung foundry (speculative), Blackwell GPU architecture (automotive-custom variant), 1,000 INT8 TOPS / 2,000 FP4 TFLOPS (4× over Orin's 254 TOPS), ARM Neoverse V3AE (v9.2-A) CPU, 64 GB LPDDR5X 4266 MHz, 273 GB/s memory bandwidth, 256 GB UFS, 350W system power. The world's first automotive SoC to integrate an Inference Transformer Engine (9× acceleration for Transformer DNN inference), supporting FP4/FP8/INT8/FP16 quantization. DriveOS 7 + Halos safety system, ISO 26262 ASIL-D + ISO 21434 dual certification.
Core Specifications
| Item | Spec |
|---|---|
| Architecture | Blackwell GPU architecture (automotive-custom, no HBM) |
| Process | 4nm Samsung 4LPP (speculative) |
| GPU Compute | 1,000 INT8 TOPS / 2,000 FP4 TFLOPS (4× Orin's 254 TOPS) |
| GPU Compute | 1,000 FP8 TFLOPS (Transformer-optimized) |
| FP4 Compute | 2,000 FP4 TFLOPS (quantization-aware training + deployment) |
| CPU | ARM Neoverse V3AE (v9.2-A, SMP) 14 cores (speculative) |
| CPU Performance | 2.3× Orin (Cortex-A78A) SPECrate 2017_int_base |
| Integrated Neural Engine | Inference Transformer Engine (first automotive SoC integration) |
| Transformer Acceleration | 9× DNN inference (vs without Transformer Engine) |
| Memory | 64 GB LPDDR5X @ 4266 MHz |
| Memory Bandwidth | 273 GB/s (256-bit LPDDR5X) |
| Storage | 256 GB UFS |
| PVA Vision Accelerator | 512 INT16 GMAC/s (Orin: 2048 INT8 GMAC/s × 2) |
| ISP | 3.5 Gigapixels/s (1.9× Orin's 1.85 GP/s) |
| Video Encode | 3.1 Gigapixels/s H.265 |
| Video Decode | 2.9 Gigapixels/s H.265 |
| Camera Interfaces | 16× GMSL 2 + 2× GMSL 3 (18 total) |
| Ethernet | 76 Gb/s (for LiDAR / radar) |
| Automotive Interfaces | 4× CAN, PCIe Gen 5 |
| Safety MCU | Renesas U2A16 (ASIL-D independent monitor) |
| System Power | 350W (SKU10/SKU12 tiers) |
| Operating Temperature | 0-35°C (SKU10) / 0-45°C (SKU12) |
| Input Voltage | 9-16V (static) / 7-32V (transient) |
| Dimensions | DevKit 15×37×15 cm / 6.2 kg |
| First OEM | ZEEKR (2025 H2 production), Volvo EX90 (2025-2026) |
| Launch Date | 2024 production, 2025-08-25 DevKit pre-sale, 2025-09 DevKit shipping |
| DevKit Price | $5,499 (DRIVE AGX Thor Developer Kit) |
Comparison with Orin
| Metric | DRIVE Thor | DRIVE Orin | Improvement |
|---|---|---|---|
| Architecture | Blackwell | Ampere | Next-gen |
| Process | 4nm | 8nm Samsung | 2× transistor density |
| GPU Compute (INT8) | 1,000 TOPS | 254 TOPS | 4× |
| GPU Compute (FP4) | 2,000 TFLOPS | N/A | New |
| Transformer Acceleration | 9× (dedicated engine) | None | New |
| CPU | Neoverse V3AE | Cortex-A78A | 2.3× SPECrate |
| Memory | 64 GB LPDDR5X | 32 GB LPDDR5 | 2× capacity |
| Memory Bandwidth | 273 GB/s | 204.8 GB/s | +33% |
| ISP | 3.5 GP/s | 1.85 GP/s | 1.9× |
| Video Encode | 3.1 GP/s | 1.0 GP/s | 3.1× |
| System Power | 350W | 60W | 5.8× |
| Cameras | 16× GMSL 2 + 2× GMSL 3 | 16× GMSL 2 | +2× GMSL 3 |
| Safety MCU | Renesas U2A16 | Integrated | Independent monitor |
| Software Stack | DriveOS 7 + Halos | DriveOS 6 | Next-gen safety |
| Certification | ASIL-D + ISO 21434 | ASIL-B | Highest safety |
Use Cases
- L2+ to L4 autonomous driving (Tesla FSD, Mercedes Drive Pilot, Volvo XC90)
- Cockpit-domain fusion (cluster + infotainment + ADAS + parking on a single SoC)
- In-vehicle large model inference (LLM, VLM, on-device GPT for in-car)
- Generative AI in-vehicle (DriveAGX + Cosmos simulation training)
- Robotics / industrial automation (DRIVE Thor + Isaac Sim)
- Multi-SoC coordination (DRIVE Hyperion 9 platform: 2× Thor = 254 TOPS×2 → 508 TOPS → 2000 TOPS)
Vendor Information
| Item | Detail |
|---|---|
| Vendor | NVIDIA Corporation |
| Design | NVIDIA Santa Clara, USA (automotive-custom) |
| Foundry | Samsung South Korea (4nm 4LPP) |
| Software Stack | DriveOS 7 (safety RTOS), DriveWorks SDK, Halos full-stack safety |
| AI Frameworks | TensorRT, CUDA, Triton Inference Server, TensorRT-LLM |
| First Customer | ZEEKR (2025 H2 production vehicles), Volvo EX90 (2026), Polestar, Mercedes-Benz |
| DevKit Price | $5,499 (shipping 2025-09) |
| Automotive Certification | ISO 26262 ASIL-D, ISO/SAE 21434 (cybersecurity) |
| Hyperion 9 Platform | Dual Thor SoC + DriveOS + 14 cameras + 9 radars + 1 LiDAR + 12 ultrasonic |
Key Features
- Blackwell GPU architecture (automotive-custom variant)
- 1,000 INT8 TOPS (4× Orin's 254 TOPS)
- 2,000 FP4 TFLOPS (first automotive FP4 support)
- Inference Transformer Engine (first automotive Transformer acceleration, 9× speedup)
- Neoverse V3AE CPU (ARM v9.2-A, 2.3× Orin CPU)
- 64 GB LPDDR5X 273 GB/s (largest memory in automotive SoC)
- 18 camera channels (16× GMSL 2 + 2× GMSL 3)
- 76 Gb/s Ethernet (10× typical automotive)
- ISO 26262 ASIL-D (highest automotive functional safety level)
- ISO/SAE 21434 (automotive cybersecurity)
- DriveOS 7 safety RTOS + Halos full-stack safety system
- Multi-Instance GPU (MIG) support (IVI + ADAS domain isolation)
Related Cards
- NVIDIA DRIVE Orin (254 TOPS Autonomous Driving) — Previous-gen automotive
- NVIDIA Jetson Thor (Robotics SoC) — Same-gen robotics
- NVIDIA H100 / B200 (Data Center) — Same-gen data center
- NVIDIA RTX 5090 (Consumer GPU) — Consumer
- Mobileye EyeQ6 / Qualcomm Snapdragon Ride — Competing automotive
- Tesla AI Training Chip Dojo V1/V2 — Training counterpart
- NVIDIA Halos Safety System
- NVIDIA DRIVE AGX Thor Developer Kit