Jetson Thor
Last reviewed
Sources
18 citations
Review status
Source-backed
Revision
v5 ยท 4,777 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
18 citations
Review status
Source-backed
Revision
v5 ยท 4,777 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA Jetson Thor (also marketed as Jetson AGX Thor) is a high-performance edge AI computing module series developed by NVIDIA for robotics, humanoid systems, and "physical AI" workloads at the edge. Built around the Blackwell architecture GPU, a 14-core Arm Neoverse-V3AE CPU, and up to 128 GB of LPDDR5X memory, the platform was previewed by NVIDIA CEO Jensen Huang at GTC 2024 alongside Project GR00T and reached general availability on August 25, 2025 with a developer kit starting at US$3,499. [1] The flagship Jetson T5000 module delivers up to 2,070 FP4 sparse TFLOPS of AI compute inside a configurable 40 to 130 watt power envelope, which NVIDIA reports as 7.5 times the AI performance and 3.5 times the energy efficiency of the previous Jetson Orin AGX generation. [1][2] Jetson Thor sits at the top of the Jetson product family and is intended to bring data center class generative AI workloads, including large language models, vision language models, and vision language action policies, onto autonomous machines such as humanoid robots, autonomous mobile robots, drones, surgical robots, and inspection systems. [2]
The Jetson Thor series belongs to NVIDIA's Jetson family of embedded AI computing platforms, which has been used since 2014 to bring GPU acceleration to autonomous machines, edge servers, and research robotics. By the mid 2020s, the rise of large multimodal models and generalist humanoid robotics created a need for far more on-device compute than the Ampere based Jetson Orin family could provide. NVIDIA framed this transition as the "physical AI" era, in which robots reason about the physical world using foundation models rather than purely deterministic perception and control stacks. [2]
Thor was first unveiled in March 2024 at GTC 2024 in San Jose. During his keynote, Jensen Huang introduced Project GR00T, a general purpose foundation model for humanoid robots, and announced that Thor would serve as the on-robot compute platform for next generation humanoids. [5] NVIDIA initially described Thor as a system on a chip incorporating a next generation Blackwell GPU with a transformer engine, an integrated functional safety processor, and 100 Gb networking, delivering 800 teraflops of 8-bit floating point performance. [5] Over the following 17 months NVIDIA refined the product into two module variants and a developer kit, with revised peak figures based on the FP4 sparse format introduced by Blackwell.
General availability arrived on August 25, 2025, when NVIDIA announced that the Jetson AGX Thor Developer Kit and Jetson T5000 production module were shipping through authorized distributors worldwide. [1] The launch was positioned as the foundation for what NVIDIA called "the age of general robotics," with early adopters spanning humanoid developers, industrial OEMs, surveying firms, and medical device makers. [1] A second module, the Jetson T4000, was disclosed for release in 2026 at a lower price and reduced power envelope. [1]
| Date | Event |
|---|---|
| March 18, 2024 | Jensen Huang previews Jetson Thor and Project GR00T at GTC 2024; initial spec cited 800 FP8 TFLOPS |
| March 2025 | NVIDIA confirms timing and additional robotics partners at GTC 2025 |
| August 25, 2025 | General availability announced; Jetson AGX Thor Developer Kit at $3,499 and Jetson T5000 module shipping |
| Q3 2025 | First developer kits delivered to early adopters and resellers |
| September 2025 | DRIVE AGX Thor developer kits (automotive sibling) begin shipping |
| October 2025 | CoRL 2025 demonstrations of GR00T N1.5 and N1.6 running on Jetson Thor |
| 2026 | Jetson T4000 module released for cost sensitive applications |
Jetson Thor is implemented as a single 87 by 100 mm system on module (SoM) with a 699-pin board-to-board connector. [3] The SoM is not pin compatible with Jetson AGX Orin because Thor uses a different connector, a higher peak power budget (up to 130 W versus 60 W for AGX Orin), and additional high speed interfaces such as PCIe Gen5 and 25 GbE. [4] NVIDIA publishes a dedicated Orin to Thor migration application note (DA-11926-001) to help integrators move existing carriers to the new platform. [4]
The GPU is based on NVIDIA's Blackwell architecture, the same family used in data center products such as the B100 and B200 but configured for an embedded power envelope. [2] Key GPU traits on the T5000 module include:
At the T4000 tier the GPU is cut down to 1,536 CUDA cores, 64 Tensor Cores, and 6 TPCs, with a lower 1.53 GHz peak clock. [3]
NVIDIA publishes peak AI throughput at multiple precisions. The numbers below are taken from the Jetson Thor Series Modules datasheet (DS-11945-001) and the Jetson Thor launch technical blog. [2][3]
| Precision | T5000 peak | T4000 peak |
|---|---|---|
| FP4 sparse | 2,070 TFLOPS | 1,200 TFLOPS |
| FP4 dense or FP8 sparse or INT8 sparse | 1,035 TFLOPS or TOPS | preliminary |
| FP8 dense or FP16 sparse | 517 TFLOPS | preliminary |
| FP16 dense | 258 TFLOPS | preliminary |
| FP32 | 8.064 TFLOPS | preliminary |
| DLA INT8 (PVA assisted vision pipeline) | up to 105 TOPS | preliminary |
The 7.5 times improvement over Jetson AGX Orin that NVIDIA quotes is calculated using FP4 sparse on Thor against INT8 sparse on Orin, the highest published precision for each generation respectively. [2]
Thor switches from the Arm Cortex-A78AE used in Orin to the Arm Neoverse-V3AE, an automotive enhanced variant of the Neoverse V3 server class core. [8] The T5000 ships with 14 Neoverse-V3AE cores; the T4000 has 12. [3] Each core has 1 MB of L2 cache and the cluster shares 16 MB of L3 cache, with a peak frequency of 2.6 GHz. [3] The use of Neoverse means that JetPack 7 software aligns with the Arm Server Base System Architecture (SBSA), the same firmware and interface specification used on Arm based server silicon. [8] This alignment lets the same CUDA 13 binaries run on Arm servers and on Jetson Thor without recompilation, which is a deliberate move toward unifying the data center and edge developer experience. [8]
The T5000 carries 128 GB of unified LPDDR5X memory on a 256-bit bus at 4,266 MHz, delivering approximately 273 GB/s of bandwidth between the CPU, GPU, and accelerators. [3] The T4000 drops to 64 GB on the same 256-bit bus and the same bandwidth. [3] Storage is provided externally through an M.2 Key M slot for NVMe SSDs (the developer kit includes a 1 TB drive) and through USB 3.2 mass storage. The 128 GB pool is large enough to keep models up to roughly 100 billion parameters resident with INT4 or FP4 quantization, which is why NVIDIA markets Thor as bringing "frontier scale" inference to the edge. [2]
In addition to the GPU, Thor integrates the following on-die accelerators:
Unlike Jetson AGX Orin, which paired the GPU with two dedicated NVIDIA Deep Learning Accelerators (NVDLA v2), the Jetson Thor SoC moves the heavy lifting onto the much larger Blackwell GPU plus the PVA. [4] NVIDIA's developer documentation treats the PVA and Tensor Cores as the primary inference paths on Thor, and the public datasheet lists deep learning acceleration in terms of GPU TOPS and a separate PVA pipeline rather than a discrete NVDLA block. [3] Developers migrating NVDLA workloads from Orin are directed to run them on the GPU or PVA on Thor. [4]
Networking is a central upgrade. The T5000 exposes four 25 GbE interfaces (aggregated via a QSFP28 connector for 100 Gb total) plus a 5 GbE RJ45 port, while the developer kit adds Wi-Fi 6E and Bluetooth via an M.2 Key E module. [11] PCIe Gen5 lanes are configurable up to x8 plus x4 plus x2, supporting NVMe SSDs, accelerator cards, and FPGA peripherals. [3]
Industrial interfaces include CAN bus, multiple UARTs, SPI, I2C, I2S, and dozens of GPIO pins. The platform also introduces the Holoscan Sensor Bridge, an Ethernet based time synchronized sensor input path originally developed for the Holoscan medical imaging SDK. [13] Sensor Bridge lets cameras and lidar units stream raw pixel and point cloud data over 10 Gb or 25 Gb Ethernet with low and deterministic latency, decoupling the sensor head from the compute module and enabling longer cable runs than MIPI CSI-2 supports. [13]
The Jetson Thor family launched with two production modules and one developer kit. The T5000 is the flagship; the T4000 is a lower cost and lower power option targeted at workloads that do not need full Blackwell scale. [3]
| Component | Jetson T5000 | Jetson T4000 |
|---|---|---|
| GPU | 2,560 CUDA cores, 96 Tensor Cores, 10 TPCs, MIG | 1,536 CUDA cores, 64 Tensor Cores, 6 TPCs, MIG |
| GPU clock | 1.57 GHz | 1.53 GHz |
| FP4 sparse peak | 2,070 TFLOPS | 1,200 TFLOPS |
| CPU | 14-core Arm Neoverse-V3AE @ 2.6 GHz | 12-core Arm Neoverse-V3AE |
| Memory | 128 GB LPDDR5X, 256-bit, 273 GB/s | 64 GB LPDDR5X, 256-bit, 273 GB/s |
| Power profile | 40 to 130 W configurable | 40 to 70 W configurable |
| 25 GbE links | 4 (via QSFP28) | 3 |
| Video encoders | 2 NVENC | 1 NVENC |
| Form factor | 87 by 100 mm SoM, 699-pin | 87 by 100 mm SoM, 699-pin |
| Module SKU | 900-13834-0080-000 | preliminary |
| Module price (1K units) | $2,999 | $1,999 |
The Jetson AGX Thor Developer Kit packages a Jetson T5000 module on a reference carrier with active cooling, a 1 TB NVMe SSD in an M.2 Key M slot, a Wi-Fi 6E plus Bluetooth M.2 Key E module, HDMI 2.0b, DisplayPort 1.4a, 5 GbE, four 25 GbE via QSFP28, multiple USB 3.2 Type-A and Type-C ports, CAN, UART, I2S audio headers, JTAG, and a 140 W power adapter. [11] The assembled kit measures roughly 243 by 112 by 57 mm. NVIDIA published the developer kit price at US$3,499, with preorders opening in August 2025 and shipments starting in Q3 2025 and continuing into November 2025 through resellers such as Arrow, RS Components, and Seeed Studio. [1][11]
Thor is the direct successor to the Jetson AGX Orin family, which itself replaced Jetson AGX Xavier in 2022. The table below summarizes the principal differences between the highest end Orin and Thor SKUs. [16]
| Feature | Jetson AGX Orin 64 GB | Jetson AGX Thor (T5000) |
|---|---|---|
| GPU architecture | Ampere | Blackwell |
| CUDA cores | 2,048 | 2,560 |
| Tensor Cores | 64 (3rd gen) | 96 (5th gen) |
| Peak AI throughput | 275 INT8 sparse TOPS | 2,070 FP4 sparse TFLOPS |
| Transformer engine | No | Yes (FP4 and FP8) |
| Multi-Instance GPU | No | Yes |
| Deep learning accelerator | 2x NVDLA v2 | None (workload moves to GPU and PVA) |
| CPU | 12-core Arm Cortex-A78AE | 14-core Arm Neoverse-V3AE |
| CPU peak clock | ~2.2 GHz | 2.6 GHz |
| Memory | 64 GB LPDDR5 at 204.8 GB/s | 128 GB LPDDR5X at 273 GB/s |
| Networking | 10 GbE | 4 by 25 GbE plus 5 GbE |
| PCIe | Gen4 | Gen5 |
| Power envelope | 15 to 60 W | 40 to 130 W |
| Form factor | SoM, 699-pin | SoM, 699-pin (not pin compatible) |
| Pin compatibility | n/a | Requires new carrier |
NVIDIA reports public benchmark speedups for generative AI workloads on the Jetson Thor launch blog. The figures below compare Thor against AGX Orin 64 GB using the latest available kernels at launch. [2]
| Benchmark | Speedup on Thor |
|---|---|
| Llama 3.3 70B inference | 1.71x |
| Qwen3-32B inference | 4.70x |
| DeepSeek-R1-Distill-Qwen-32B | 4.87x |
| Qwen2.5-VL-7B with FP4 plus speculative decoding | 3.5x |
| GR00T N1.5 vision language action model | 2.74x |
| Aggregate generative AI inference | up to 5x |
In multimodal serving tests with Qwen2.5-VL-3B and Llama 3.2 3B handling 16 simultaneous requests, NVIDIA reported time to first token below 200 milliseconds and time per output token below 50 milliseconds, which it positions as the threshold for natural human and robot interaction. [2]
Thor uses JetPack 7.0 (and later releases such as 7.2 with CUDA 13.2) as its base software stack. [7] JetPack 7 is built on the Linux kernel 6.8 and Ubuntu 24.04 LTS, includes a preemptable real time patch set for deterministic execution, and aligns Jetson firmware and bootloaders with the Arm SBSA specification used on Arm based servers. [7][12] The result is that the same Arm CUDA 13.0 binaries used on data center Arm Grace systems will run on Jetson Thor, which simplifies sim to real workflows. [8]
Key software components for Jetson Thor include:
On the model and robotics side, NVIDIA pairs Jetson Thor with several platforms in its Physical AI stack:
Jetson Thor supports the major open ecosystem frameworks, including PyTorch, TensorFlow, ONNX Runtime, vLLM, Hugging Face Transformers, llama.cpp, and the Triton inference server, plus model families from Meta (Llama 3 and 4), Alibaba (Qwen 2.5 and Qwen 3), DeepSeek, Google Gemma, and Mistral. [2]
| Metric | Target performance |
|---|---|
| Time to first token (TTFT) | under 200 ms for 7B class multimodal models |
| Time per output token (TPOT) | under 50 ms for 7B class multimodal models |
| Generative AI speedup vs Orin | up to 5x aggregate; 7.5x for FP4 quantized inference |
| Speculative decoding speedup | up to 2x additional |
Thor exposes a wide range of power and thermal modes that integrators can pick using the nvpmodel tool. NVIDIA describes the envelope from low power deployment to peak compute, with corresponding clock and core configurations. [3]
| Profile (T5000) | Approximate power | Target use case |
|---|---|---|
| Low | 40 W | Battery operated robots, drones, evaluation on passive cooling |
| Mid | 60 to 80 W | Standard humanoid and AMR workloads, sustained inference |
| Performance | 100 W | Multi-model multimodal serving, generative AI with FP4 |
| Maximum | 130 W | Peak burst performance for VLA, LLM, and large VLM workloads |
The T4000 caps at roughly 70 W, suiting passively cooled and battery powered systems where the smaller GPU and memory pool are sufficient. [3]
NVIDIA positions Jetson Thor as a single platform spanning several robotics and edge AI segments. Highlighted applications include: [1]
A recurring theme in NVIDIA's launch material is the ability to run multiple AI models concurrently on one device. Thor can simultaneously execute a vision language action policy, a higher level reasoning model, a perception network, and a safety classifier through MIG partitions, an architecture NVIDIA refers to as a "multi-model robot brain." [18]
NVIDIA disclosed a broad list of early Jetson Thor adopters at GA. The roster spans humanoid developers, industrial OEMs, surveying and inspection firms, and medical device makers. [1]
| Adopter | Sector | Application |
|---|---|---|
| Agility Robotics | Humanoid robots | Sixth generation Digit humanoid for logistics |
| Amazon Robotics | Warehouse robotics | Internal robotics fleet for fulfillment |
| Apptronik | Humanoid robots | Apollo humanoid platform |
| Boston Dynamics | Legged and humanoid robots | Atlas humanoid, expanded GR00T integration |
| Caterpillar | Heavy equipment | Autonomous construction and mining vehicles |
| Figure | Humanoid robots | Figure 02 and successors |
| Hexagon | Surveying and metrology | Reality capture and inspection robots |
| Medtronic | Medical robotics | Surgical and imaging systems |
| Meta | AI research | Physical AI and humanoid research |
| 1X Technologies | Humanoid robots | EVE and NEO platforms (evaluating) |
| John Deere | Agriculture | Autonomous farm equipment (evaluating) |
| OpenAI | AI research | Robotics research (evaluating) |
| Physical Intelligence | Robotics foundation models | VLA model deployment (evaluating) |
| Fourier Intelligence | Humanoid robots | GR1 and successors via GR00T |
| Sanctuary AI | Humanoid robots | Phoenix humanoid via GR00T |
| Unitree Robotics | Quadrupeds and humanoids | H1 and G1 platforms via GR00T |
| XPENG Robotics | Humanoid robots | Iron humanoid via GR00T |
In aggregate, NVIDIA reported more than 7,000 customers deploying Jetson Orin hardware by 2025, over 150 hardware partners offering production ready solutions, and more than 2 million developers active on its robotics platforms. [1]
NVIDIA uses the Thor brand for two distinct but related products. Jetson Thor is the robotics and physical AI variant, while DRIVE AGX Thor is the automotive variant introduced for advanced driver assistance and autonomous driving. [8] Both products share the underlying Thor SoC: a Blackwell GPU with the Transformer Engine, Arm Neoverse-V3AE CPU cores, the same memory subsystem, and similar peak compute around 2,000 FP4 TFLOPS. [8] They diverge on packaging and software:
NVIDIA's product positioning is that vehicles and robots increasingly share the same software stack, so unifying compute across the two domains lowers cost for OEMs that build both autonomous vehicles and robots. [8]
Jetson Thor competes with a range of edge and embedded accelerators, though no other vendor matches its combination of Blackwell class GPU, 128 GB memory, and integrated Arm server class CPU. The table below sketches the broader market context as of late 2025 and early 2026. [17]
| Platform | Peak AI throughput | Memory | Power | Notable traits |
|---|---|---|---|---|
| NVIDIA Jetson AGX Thor (T5000) | 2,070 FP4 sparse TFLOPS | 128 GB LPDDR5X | 40 to 130 W | Full Blackwell GPU plus Arm Neoverse, JetPack 7, Isaac and GR00T |
| NVIDIA Jetson AGX Orin 64 GB | 275 INT8 sparse TOPS | 64 GB LPDDR5 | 15 to 60 W | Ampere GPU, NVDLA, mature ecosystem |
| Qualcomm Robotics RB6 | ~200 INT8 TOPS (combined) | up to 16 GB LPDDR5 | ~30 W | 5G connectivity, Adreno GPU, lower compute density |
| Qualcomm Dragonwing IQ10 | tens of INT8 TOPS | LPDDR5 | low watts | Power efficient alternative for non-frontier robotics |
| Hailo-15 | 20 TOPS at sub-3 W | shared host memory | ~3 W | Dedicated dataflow accelerator, low power vision |
| Hailo-10 | up to 40 INT8 TOPS | host memory | ~3.5 W | Designed for client devices and small robots |
| Mythic M1076 | up to 25 TOPS analog | analog in-memory | ~3 to 4 W | Analog compute in memory, niche deployments |
| Google Edge TPU (Coral) | 4 INT8 TOPS | host memory | ~2 W | Small models only, very low power |
| AMD Versal AI Edge | tens of TOPS | DDR4 or LPDDR4 | ~30 W | FPGA fabric plus AI engines, deterministic latency |
In practice, Thor competes most directly with high end Jetson Orin and DRIVE platforms rather than with low power accelerators like Hailo or Coral. The closer comparisons are NVIDIA's own AGX Orin Industrial and the broader Arm server market, where Thor's SBSA alignment means workloads can move between Grace based servers and the edge module without code changes. [8]
NVIDIA provides a 10-year production lifecycle for Jetson Thor modules, matching the long support windows expected by industrial and automotive integrators. [3] Secure boot, A or B firmware updates, and a hardware root of trust are standard. JetPack 7 introduces a stronger SBSA-aligned firmware model, with UEFI and ACPI tables on the SoC, which simplifies long term operating system support and field updates. [7]
For developers, NVIDIA offers a remote provisioning workflow using NVIDIA Mission Control style tools, a containerized application layer based on the NVIDIA Container Toolkit, and prebuilt model containers in the NGC registry. The Jetson Thor developer kit boots into a desktop Ubuntu environment with CUDA, TensorRT, and Isaac libraries preinstalled, and is intended to act as a "robot brain" out of the box. [10]
Industry coverage at launch focused on three themes. First, Thor's headline 7.5 times AI compute improvement over Jetson AGX Orin was widely cited as a generational leap rather than an incremental upgrade, although several reviewers noted that the comparison mixes FP4 sparse on Thor against INT8 sparse on Orin. [10] Second, the introduction of MIG to an embedded platform was seen as a significant step toward multi model robotics, where perception, planning, language reasoning, and safety can coexist on a single SoC. [18] Third, the price of US$3,499 for the developer kit and US$2,999 for the bulk module was noted as high by hobbyist standards but competitive against discrete GPU plus host CPU solutions targeted at humanoid robots. [11]
Reviews from HotHardware and ServeTheHome highlighted hands-on demonstrations of GR00T N1 running inference on Thor while controlling a simulated robot performing precision tasks, with NVIDIA generating over 750,000 simulated trajectories in roughly 11 hours and reporting 40 percent performance improvements from synthetic data augmentation. [10] CNX Software and the Edge AI and Vision Alliance focused on the new networking stack (4 by 25 GbE plus 5 GbE plus Wi-Fi 6E) and the Holoscan Sensor Bridge as a clean fit for camera and lidar heavy robots. [11][13]