Skip to main content

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

ItemSpec
ArchitectureBlackwell GPU architecture (automotive-custom, no HBM)
Process4nm Samsung 4LPP (speculative)
GPU Compute1,000 INT8 TOPS / 2,000 FP4 TFLOPS ( Orin's 254 TOPS)
GPU Compute1,000 FP8 TFLOPS (Transformer-optimized)
FP4 Compute2,000 FP4 TFLOPS (quantization-aware training + deployment)
CPUARM Neoverse V3AE (v9.2-A, SMP) 14 cores (speculative)
CPU Performance2.3× Orin (Cortex-A78A) SPECrate 2017_int_base
Integrated Neural EngineInference Transformer Engine (first automotive SoC integration)
Transformer Acceleration DNN inference (vs without Transformer Engine)
Memory64 GB LPDDR5X @ 4266 MHz
Memory Bandwidth273 GB/s (256-bit LPDDR5X)
Storage256 GB UFS
PVA Vision Accelerator512 INT16 GMAC/s (Orin: 2048 INT8 GMAC/s × 2)
ISP3.5 Gigapixels/s (1.9× Orin's 1.85 GP/s)
Video Encode3.1 Gigapixels/s H.265
Video Decode2.9 Gigapixels/s H.265
Camera Interfaces16× GMSL 2 + 2× GMSL 3 (18 total)
Ethernet76 Gb/s (for LiDAR / radar)
Automotive Interfaces4× CAN, PCIe Gen 5
Safety MCURenesas U2A16 (ASIL-D independent monitor)
System Power350W (SKU10/SKU12 tiers)
Operating Temperature0-35°C (SKU10) / 0-45°C (SKU12)
Input Voltage9-16V (static) / 7-32V (transient)
DimensionsDevKit 15×37×15 cm / 6.2 kg
First OEMZEEKR (2025 H2 production), Volvo EX90 (2025-2026)
Launch Date2024 production, 2025-08-25 DevKit pre-sale, 2025-09 DevKit shipping
DevKit Price$5,499 (DRIVE AGX Thor Developer Kit)

Comparison with Orin

MetricDRIVE ThorDRIVE OrinImprovement
ArchitectureBlackwellAmpereNext-gen
Process4nm8nm Samsung2× transistor density
GPU Compute (INT8)1,000 TOPS254 TOPS
GPU Compute (FP4)2,000 TFLOPSN/ANew
Transformer Acceleration (dedicated engine)NoneNew
CPUNeoverse V3AECortex-A78A2.3× SPECrate
Memory64 GB LPDDR5X32 GB LPDDR52× capacity
Memory Bandwidth273 GB/s204.8 GB/s+33%
ISP3.5 GP/s1.85 GP/s1.9×
Video Encode3.1 GP/s1.0 GP/s3.1×
System Power350W60W5.8×
Cameras16× GMSL 2 + 2× GMSL 316× GMSL 2+2× GMSL 3
Safety MCURenesas U2A16IntegratedIndependent monitor
Software StackDriveOS 7 + HalosDriveOS 6Next-gen safety
CertificationASIL-D + ISO 21434ASIL-BHighest 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

ItemDetail
VendorNVIDIA Corporation
DesignNVIDIA Santa Clara, USA (automotive-custom)
FoundrySamsung South Korea (4nm 4LPP)
Software StackDriveOS 7 (safety RTOS), DriveWorks SDK, Halos full-stack safety
AI FrameworksTensorRT, CUDA, Triton Inference Server, TensorRT-LLM
First CustomerZEEKR (2025 H2 production vehicles), Volvo EX90 (2026), Polestar, Mercedes-Benz
DevKit Price$5,499 (shipping 2025-09)
Automotive CertificationISO 26262 ASIL-D, ISO/SAE 21434 (cybersecurity)
Hyperion 9 PlatformDual 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)