# NVIDIA Jetson T4000

> Source: https://aiwiki.ai/wiki/jetson_t4000
> Updated: 2026-07-24
> Categories: AI Hardware, NVIDIA, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**NVIDIA Jetson T4000** is a commercial system-on-module developed by [Nvidia](/wiki/nvidia) for [edge AI](/wiki/edge_ai) and [robotics](/wiki/robotics). It is a member of the [NVIDIA Jetson](/wiki/nvidia_jetson) Thor family and became available on January 5, 2026. The module combines a [NVIDIA Blackwell](/wiki/nvidia_blackwell) graphics processing unit with a 12-core Arm CPU and 64 GB of LPDDR5X memory.[1][2]

T4000 is a production module rather than a complete computer or developer kit. It requires a compatible carrier board, external storage, power delivery, cooling, and a flashed software image. The Jetson AGX Thor Developer Kit uses the higher-tier T5000 module, not T4000. Physical Ethernet, USB, display, camera, and other connectors depend on the carrier board.[2][3]

NVIDIA rates T4000 at up to 1,200 FP4 TFLOPS using structured sparsity. This is a theoretical low-precision Tensor Core peak, not measured application throughput. The module's current default power profile is 70 W, while NVIDIA markets a configurable 40 W to 70 W envelope. Current documentation also provides an experimental MAXN mode without a defined power budget.[3][4]

| Field | NVIDIA Jetson T4000 |
| --- | --- |
| Product type | Commercial system-on-module |
| Family | Jetson Thor |
| Commercial module | P3834-0000 |
| Availability | January 5, 2026 |
| GPU | Blackwell, 1,536 CUDA cores, 6 TPCs |
| CPU | 12 Arm Neoverse-V3AE cores |
| Memory | 64 GB LPDDR5X, 273 GB/s |
| Sparse FP4 peak | 1,200 TFLOPS |
| Default power profile | 70 W |
| Marketed configurable range | 40 W to 70 W |
| Module dimensions | 87.0 x 100 x 15.29 mm |
| Current suggested price | $2,999 each at 1,000-unit volume |
| Stated availability horizon | January 2036 |

## Product identity and availability

NVIDIA announced that T4000 was available during CES on January 5, 2026. Its developer documentation identifies the commercial product as `Jetson T4000 64GB (P3834-0000)`. The launch should not be confused with the 2025 introduction of the broader [Jetson Thor](/wiki/jetson_thor) platform or with availability of the T5000-based developer kit.[1][2]

The module measures 87.0 x 100 x 15.29 mm and uses a 699-pin board-to-board connector. It includes a thermal transfer plate with a heatpipe, but not a complete thermal solution. T4000 and T5000 have the same form factor and are pin-compatible. A carrier or cooling design that accepts both modules still has to account for their different power limits, I/O features, memory capacity, and thermal behavior.[3]

NVIDIA lists T4000 availability through January 2036. That lifecycle date indicates the company's current supply commitment, not guaranteed inventory at every distributor or a promise that software releases will support every deployed image until that date.[7]

## Processing and memory architecture

The integrated [graphics processing unit](/wiki/gpu) has two graphics processing clusters, six texture processing clusters, and 1,536 CUDA cores. It uses fifth-generation Tensor Cores and can run the GPU at up to 1.53 GHz. The datasheet lists 4.700 TFLOPS of FP32 throughput. It also documents hardware support for Multi-Instance GPU partitioning, including a possible split of four and two TPCs. Software support and qualification still depend on the installed JetPack release.[2][3]

The CPU contains twelve 64-bit Arm Neoverse-V3AE cores implementing Armv9.2-A. Each core has 64 KB instruction and 64 KB data L1 caches and a 1 MB L2 cache, with 16 MB of shared L3. The CPU can reach 2.6 GHz in MAXN, while the default 70 W profile caps it at 1.998 GHz. A PVA 3.0 accelerator provides additional processing for [computer vision](/wiki/computer_vision) workloads.[3][4]

T4000 has 64 GB of LPDDR5X memory on a 256-bit interface. NVIDIA specifies a 4,266 MHz memory clock and 273 GB/s bandwidth. The module supports memory encryption and protected regions, but the T4000 specification does not list the alternate-link ECC feature found on T5000. Applications that require end-to-end error correction should not infer T5000 memory behavior from the shared family name.[3]

The module contains 64 MB of NOR flash for platform functions. Application storage is external, using an NVMe device over up to four PCIe lanes or storage connected through USB 3.2. A commercial T4000 module does not include an NVMe drive.[3]

## I/O, cameras, and media

T4000 integrates three multigigabit Ethernet MACs, each capable of up to 25 Gbps. These are controller interfaces and module signals, not three physical network ports. A carrier board must supply compatible physical-layer devices and connectors. T4000 also exposes up to eight PCIe Gen5 lanes, along with USB, UART, SPI, I2C, I2S, PWM, audio, and display interfaces.[2][3]

Camera support includes 16 MIPI CSI-2 lanes, up to six active cameras, and up to 32 virtual channels using C-PHY or D-PHY. NVIDIA also specifies support for up to 20 cameras through the Holoscan Sensor Bridge. T4000 has one image signal processor, compared with two on T5000.[2][3]

The module contains one NVENC and one NVDEC engine. It can encode H.264 and H.265 and decode H.265, H.264, VP9, VP8, AV1, MPEG-4, MPEG-2, and VC-1. Four shared HDMI 2.1 or DisplayPort 1.4a HBR3 MST outputs can drive resolutions up to 7680 x 4320 at 30 Hz. Actual connectors and concurrently usable combinations depend on the carrier and system design.[3]

No native CAN controller is listed for T4000. Some partner computers advertise CAN, Power over Ethernet, extra storage, or different network speeds because their carrier boards add those functions. Those system features are not properties of the module itself.[3]

## AI throughput and benchmark context

The 1,200 TFLOPS headline uses sparse FP4 operations. NVIDIA's datasheet also lists up to 600 sparse FP8 TFLOPS, 600 dense FP4 TFLOPS, and 300 dense FP8 TFLOPS. Structured sparsity can double the reported operation rate by skipping selected zero values. FP4 and FP8 results therefore should not be compared directly with older INT8 TOPS or FP16 figures without matching precision, sparsity, and operation-count conventions.[3]

NVIDIA reported 218 tokens per second for Qwen3-30B-A3B, 68 for Qwen3 32B, 40 for Nemotron 12B, 64 for DeepSeek R1 Distill Qwen 32B, and 100 for Mistral 3 14B on T4000. These results show the vendor's intended use for on-device [large language model](/wiki/large_language_model) inference. The published table does not fully disclose batch size, prompt and output lengths, quantization, or memory settings, so the numbers are not reproducible latency guarantees.[2]

NVIDIA also promotes T4000 for [vision language model](/wiki/vision_language_model) and [vision-language-action model](/wiki/vision_language_action_model) pipelines and claims up to twice the performance of Jetson AGX Orin in its launch material. The public page does not provide enough T4000-specific methodology for an independent reproduction of that comparison. Performance in a deployed robot depends on model architecture, precision, memory use, sensor processing, thermal design, and how workloads share the CPU, GPU, PVA, and media engines.[2]

## Power and thermal behavior

Current Jetson Linux documentation ships two predefined T4000 profiles. The default `70W` mode enables all twelve CPU cores, all six GPU TPCs, a 1.998 GHz CPU maximum, a 1.53 GHz GPU maximum, and the full 4,266 MHz memory clock. `MAXN` raises the CPU maximum to 2.601 GHz and the video-engine maximum from 1.557 to 1.692 GHz, but it has no defined module power budget.[4]

The marketed 40 W lower bound is not a named predefined profile in the current power-mode table. NVIDIA provides tools for generating custom modes, so a system can be configured below 70 W, but its clocks and workload performance then depend on that custom configuration. The total module overcurrent limit is 90 W, which is a protection threshold rather than a supported continuous operating target.[4]

NVIDIA describes MAXN as experimental and does not recommend prolonged heavy use. Thermal and electrical throttling can make MAXN slower or less stable than a bounded profile. The power guide warns that relaxing current-limit controls can permanently damage the hardware.[4]

The datasheet gives a 75 C maximum thermal-transfer-plate surface temperature, a junction operating range from -25 C to 115 C, and a 109 C slowdown threshold. It does not guarantee stable air-cooled operation above 45 C ambient. These limits make enclosure airflow, heat-sink design, sensor load, and sustained workload part of system qualification, especially for mobile robots.[3]

## Software stack

T4000 launched with JetPack 7.1 and Jetson Linux 38.4. As of July 24, 2026, NVIDIA lists JetPack 7.2 and Jetson Linux 39.2 as the current supported release for T4000. The release uses Ubuntu 24.04 and Linux kernel 6.8 and includes [CUDA](/wiki/cuda) 13.2.1, cuDNN 9.20.0, [TensorRT](/wiki/tensorrt) 10.16.2, VPI 4.1.3, PVA 2.9.1, DeepStream 8.0, Holoscan 3.9.0, NVIDIA Container Toolkit 1.19, and Vulkan 1.4.[2][5]

JetPack 7.2 release material describes GPU partitioning on Jetson Thor as a technology preview and explicitly names T5000 in that feature note. Although the T4000 hardware datasheet supports partitioning, the current public release note does not establish production-ready T4000 software support. The same download page labels Isaac ROS as coming soon, so it should not be assumed to be part of the JetPack 7.2 image.[3][5]

Commercial modules ship without software. A system integrator selects a supported board configuration, flashes Jetson Linux and JetPack components, and qualifies device-tree settings, carrier interfaces, storage, cooling, and update behavior. TensorRT Edge-LLM and Video Codec SDK support were part of the T4000 launch software context.[2]

## Robotics deployment

T4000 is positioned for local perception, language, planning, and control pipelines in [physical AI](/wiki/physical_ai) systems. Its memory capacity and sensor interfaces can support applications such as a [humanoid robot](/wiki/humanoid_robot), an [autonomous mobile robot](/wiki/autonomous_mobile_robot), industrial inspection, multi-camera analytics, and low-latency human-machine interaction.[1][2]

The module is only one part of such a system. Carrier-board design determines usable sensors and external networks. Cooling determines sustained clocks. The software image determines available drivers and libraries. Motor controllers, emergency stops, isolation, power conditioning, and real-time safety functions normally sit outside the module. A vendor throughput rating does not establish end-to-end response time or safe robot behavior.

Developers comparing the 64 GB T4000 with cloud inference should also account for model fit, update bandwidth, fleet management, data residency, and failure recovery. Local processing can reduce network dependency and keep sensor data near the machine, but it transfers responsibility for image maintenance, access control, monitoring, and physical serviceability to the operator.

## Security, safety, and lifecycle

Jetson Thor secure boot begins with an on-die BootROM root and can authenticate later boot firmware and UEFI payloads. NVIDIA documents public-key and symmetric-key provisioning, up to 16 public keys with the first 15 revocable, and irreversible security-fuse programming. Incorrect fuse provisioning can make a module unusable, and secure boot does not automatically secure applications, credentials, carrier-board peripherals, or update infrastructure.[8]

A March 2026 NVIDIA bulletin identified two initrd vulnerabilities in Jetson Thor Linux 38.2 and states that they were fixed in 38.4. Current deployments should use supported patched images and track later bulletins rather than assume that hardware boot features remove the need for software updates.[9]

NVIDIA's launch announcement distinguishes IGX Thor as the product line with enterprise software and functional-safety positioning. The reviewed T4000 documents do not claim a functional-safety certification. GPU partitioning, secure boot, PCB manufacturing standards, and lifecycle commitments are useful platform properties, but none proves that an autonomous machine is safe. Safety depends on the complete system, its hazards, fallback behavior, validation, and any certification required by the application.[1]

The module datasheet states a five-year 24x7 operating lifetime, and NVIDIA's FAQ lists a three-year warranty for Jetson modules. These terms are separate from the January 2036 product-availability horizon.[3][6][7]

## Ordering, regions, and export classification

NVIDIA introduced T4000 at a suggested price of $1,999 each for orders of 1,000 units. Its current FAQ lists $2,999 at the same volume, while T5000 is $4,999 and the T5000-based developer kit is $5,499. An industry report dated July 22, 2026 described the T4000 change as a 50 percent increase. Distributor quotes, lower-volume pricing, taxes, and availability can differ.[1][6][11]

The FAQ lists `900-13834-0000-001` for the general module and `900-13834-0000-0A1` for a USA-origin version. It says general modules may originate in China, the United States, or Vietnam and lists sales regions across North America, Europe, and parts of Asia-Pacific. Arrow's current distributor page marks a T4000 listing active.[6][10]

NVIDIA and Arrow classify the module under ECCN `5A992.C`. That classification does not guarantee that every sale, export, re-export, or end use is permitted. Customers must check the exact part, origin, destination, end user, end use, distributor terms, and applicable law.[6][10]

## Comparison with Jetson T5000

T4000 is the lower-power, lower-memory member of the initial commercial Jetson Thor module pair. T5000 is the higher-tier module and is the one installed in the Jetson AGX Thor Developer Kit.[1][3]

| Field | T4000 | T5000 |
| --- | ---: | ---: |
| CUDA cores | 1,536 | 2,560 |
| GPU TPCs | 6 | 10 |
| Sparse FP4 peak | 1,200 TFLOPS | 2,070 TFLOPS |
| CPU cores | 12 | 14 |
| Memory | 64 GB | 128 GB |
| Memory bandwidth | 273 GB/s | 273 GB/s |
| Integrated 25 Gbps MGBE MACs | 3 | 4 |
| NVENC and NVDEC engines | 1 and 1 | 2 and 2 |
| Native CAN interfaces | None listed | 4 |
| Alternate-link ECC | Not listed | Listed |
| Default power profile | 70 W | 120 W |
| Total module overcurrent limit | 90 W | 130 W |
| Current 1,000-unit suggested price | $2,999 | $4,999 |

The two modules share dimensions and pin compatibility, but they are not interchangeable without engineering review. T5000 needs greater power and cooling headroom and exposes functions that T4000 lacks. T4000 can be a better fit when 64 GB is sufficient and system power is constrained, while T5000 provides more compute, memory, networking, media engines, and native CAN. Neither choice removes the need to validate application performance on the final carrier, cooling system, and software image.[3][4][6]

## References

1. NVIDIA, "NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots," January 5, 2026. https://nvidianews.nvidia.com/news/nvidia-releases-new-physical-ai-models-as-global-partners-unveil-next-generation-robots
2. NVIDIA Technical Blog, "Accelerate AI Inference for Edge and Robotics With NVIDIA Jetson T4000 and NVIDIA JetPack 7.1," January 5, 2026. https://developer.nvidia.com/blog/accelerate-ai-inference-for-edge-and-robotics-with-nvidia-jetson-t4000-and-nvidia-jetpack-7-1/
3. NVIDIA, "Jetson Thor Series Modules Data Sheet," DS-11945-001 v1.3, September 29, 2025, NVIDIA-authored PDF mirrored by Plink-AI. https://plink-ai.com/uploads/20260529/a41a5a15f520ec66964d681b5513904b.pdf
4. NVIDIA, "Platform Power and Performance: Jetson Thor," Jetson Linux Developer Guide 39.2, accessed July 24, 2026. https://docs.nvidia.com/jetson/archives/r39.2/DeveloperGuide/SD/PlatformPowerAndPerformance/JetsonThor.html
5. NVIDIA, "JetPack SDK 7.2," accessed July 24, 2026. https://developer.nvidia.com/embedded/jetpack/downloads
6. NVIDIA, "Jetson FAQ," accessed July 24, 2026. https://developer.nvidia.com/embedded/faq
7. NVIDIA, "Jetson Product Lifecycle," accessed July 24, 2026. https://developer.nvidia.com/embedded/lifecycle
8. NVIDIA, "Secure Boot," Jetson Linux Developer Guide 39.2, accessed July 24, 2026. https://docs.nvidia.com/jetson/archives/r39.2/DeveloperGuide/SD/Security/SecureBoot.html
9. NVIDIA, "Security Bulletin: NVIDIA Jetson and IGX Devices - March 2026," March 31, 2026. https://nvidia.custhelp.com/app/answers/detail/a_id/5797/~/security-bulletin%3A-nvidia-jetson-and-igx-devices---march-2026
10. Arrow Electronics, "NVIDIA 900-13834-0000-000 Jetson T4000," accessed July 24, 2026. https://www.arrow.com/en/products/900-13834-0000-000/nvidia.html
11. CNX Software, "NVIDIA increases the price of Jetson modules and devkits by up to 101%," July 22, 2026. https://www.cnx-software.com/2026/07/22/nvidia-increases-the-price-of-jetson-modules-and-devkits-by-up-to-101/