Nvidia has introduced the Jetson T3000 and T2000 modules, two new variants of its Thor architecture designed to bring AI models to robots, vision systems, and autonomous machines with less memory, power consumption, and size than the current Jetson T5000. These systems can be emulated before their commercial release, scheduled for the first quarter of 2027.
The key highlights of the new Jetson Thor modules in 30 seconds
- The Jetson T3000 offers 865 teraflops FP4, 32 GB of memory, and an eight-core ARM CPU.
- The T2000 downsizes to 400 teraflops FP4 and 16 GB.
- Nvidia estimates the T3000 to be roughly half the size and power consumption of the T5000.
- Cosmos 3 Edge enables running a 4-billion-parameter robotic model on the device.
- The modules will be available during the first quarter of 2027.
The manufacturer aims to expand the Thor family beyond humanoid robots and higher-cost industrial platforms. The Jetson T3000 targets systems that require running multimodal models and real-time sensor processing but do not need the 128 GB of memory or 2,070 teraflops FP4 of the T5000. The T2000 will occupy a lower tier, suitable for mobile robots, manipulators, visual agents, and other edge AI devices.
Nvidia has not yet disclosed prices, exact dimensions, or power consumption ranges for these two new modules. A comprehensive technical sheet with interfaces, GPU cores, compatible storage, or camera configurations is also pending. For now, its specifications should be considered preliminary and limited to the data provided in the announcement.
T3000 reduces memory and power compared to Jetson T5000
The Jetson T3000 combines a Blackwell-based GPU, an eight-core ARM Neoverse CPU, 32 GB of LPDDR5X memory, and a bandwidth of 273 GB/s. Nvidia states it offers 865 teraflops of FP4 compute and 25 Gbps Ethernet connectivity.
The company claims that the module is about half the size and consumes roughly half the power of the T5000. However, it does not provide exact measurements to support this comparison, so the power range cannot be directly inferred from the 40-130 W of the higher-end model.
| Feature | Jetson T2000 | Jetson T3000 | Jetson T4000 | Jetson T5000 |
|---|---|---|---|---|
| GPU Architecture | Blackwell | Blackwell | Blackwell | Blackwell |
| Announced AI Performance | 400 TFLOPS FP4 | 865 TFLOPS FP4 | 1,200 TFLOPS FP4 dispersed | 2,070 TFLOPS FP4 dispersed |
| CPU | Not detailed | ARM Neoverse, 8 cores | ARM Neoverse-V3AE, 12 cores | ARM Neoverse-V3AE, 14 cores |
| Memory | 16 GB | 32 GB LPDDR5X | 64 GB LPDDR5X | 128 GB LPDDR5X |
| Bandwidth | Not detailed | 273 GB/s | 273 GB/s | 273 GB/s |
| Network | Not detailed | 25 GbE | 3 × 25 GbE | 4 × 25 GbE |
| Power Consumption | Not announced | About half of T5000 | 40-70 W | 40-130 W |
| Form Factor | Not announced | About half of T5000 | 100 × 87 mm | 100 × 87 mm |
| Availability | Q1 2027 | Q1 2027 | Available | Available |
The figures for the T4000 and T5000 come from Nvidia’s current official specifications of Jetson Thor. Nvidia uses dispersed FP4 teraflops to express the maximum performance of these modules. In the T3000 and T2000 announcement, they mention FP4 teraflops without clarifying whether the numbers are calculated under exactly the same conditions.
This variability prevents treating the table as a direct comparison of actual performance. Teraflops denote the theoretical capacity for certain 4-bit mathematical operations, but by themselves, they do not predict robot speed, vision model performance, or application throughput.
Final performance depends on the precision used, model size, available memory, data movement, software optimization, and sensor load. A module with fewer teraflops may suffice when the model fits in memory and the application does not heavily use the GPU.
Nvidia claims that the T3000 achieves inference performance similar to the T5000 in multimodal workloads, including large language models, vision-and-language models, VLA systems (vision-language-action), and foundational models. However, this statement is from the manufacturer and lacks independent testing or detailed model results.
Where does each module fit within Jetson Thor?
| Module | Expected Role | Application Examples |
|---|---|---|
| Jetson T2000 | Entry-level for Thor architecture | Visual agents, mobile robots, industrial manipulators |
| Jetson T3000 | Advanced robotics with lower memory and power | Humanoids, multimodal systems, sensor processing |
| Jetson T4000 | Mid-range high-performance | Complex robotics and multi-tasking loads |
| Jetson T5000 | Maximum Thor configuration | Advanced humanoids, large models, multisensor fusion |
Memory reduction partly aims at cost efficiency. Nvidia links the T3000’s development to the possibility of reducing system prices amid rising memory costs. Moving from 128 GB to 32 GB can significantly cut costs but also limits model size and the number of simultaneous tasks.
The T2000 reduces to 16 GB and 400 teraflops FP4. Nvidia presents it as an entry point to extend Thor to systems that previously used Jetson Orin or other less powerful edge AI platforms, though it hasn’t specified which Orin models it replaces or whether they will coexist for years.
Software Agents to Compensate for Reduced Memory
Alongside the modules, Nvidia has released new capabilities for development agents to analyze and optimize memory usage, system configuration, and application deployment on Jetson devices.
These tools aim to automate tasks previously requiring manual review of models, buffers, libraries, processes, and accelerators. Nvidia states they can enable running the same application on a module with less memory, though results will vary per project.
| Company or Project | Reported Memory Savings | Achieved Changes |
|---|---|---|
| UBTech, Agile Robots, and Connect Tech | Up to 15 GB | Reducing Jetson AGX Orin from 64 GB to 32 GB |
| SandStar | Up to 4 GB | Reducing Orin NX from 16 GB to 8 GB |
| NoTraffic | 30% | More space for additional functions on Jetson TX2 NX |
| GROOVE X | Not quantified | Distributing loads across heterogeneous accelerators |
All these results are provided by Nvidia and participating companies. They do not guarantee that any application can halve its memory without performance or functionality loss.
Methods may include quantization, removing unnecessary processes, memory reuse, offloading operations to specialized accelerators, and choosing smaller model versions. Some optimizations maintain performance, while others may reduce precision, context, or capacity.
The savings become especially significant with the T2000 and T3000. Running advanced models on 16 or 32 GB will require developers to better manage memory compared to a 128 GB T5000. The raw computational capacity is less meaningful if models, context, or intermediate images cannot fit into the system.
Cosmos 3 Edge runs a 4-billion-parameter robotic model
Nvidia also announced Cosmos 3 Edge, a 4-billion-parameter version of its foundational models family. It is designed to run on Thor platforms to help physical systems interpret their environment, reason about observations, and generate actions locally without relying on cloud connectivity.
The company claims developers could adapt the model to a robot type and its sensors in about a day, though this will depend on data volume, training hardware, customization level, and device complexity.
| Software Component | Role in Robotic System |
|---|---|
| Cosmos 3 Edge | Environmental interpretation and action generation |
| Isaac | Simulation, perception, and robotics development |
| Isaac GR00T | Models and tools for humanoid robots |
| Nemotron | Open language and reasoning models |
| Jetson Agent Skills | Memory tuning, configuration, deployment |
| NemoClaw | Coordination and execution of agents |
Cosmos 3 Edge seeks to narrow the gap between simulation and the physical world. A developer can train or tailor behaviors virtually and then deploy the model for local inference on the robot.
On-device processing reduces latency and allows operation even with limited connectivity. It also avoids continually sending video and sensor data to remote centers, although protecting stored or transmitted information remains essential.
Jetson and IGX T3000 are not exactly the same product
Nvidia is preparing two versions of the T3000. Jetson T3000 targets robotics and general-purpose embedded systems, while IGX T3000 incorporates the same computational power but adds features for functional safety and compatibility with Nvidia Halos for Robotics.
| Platform | Focus | Main Difference |
|---|---|---|
| Jetson T3000 | Robots and autonomous machines | Integrated AI and processing module |
| IGX T3000 | Industrial environments and human-interacting robots | Built-in functional safety features |
| Jetson T2000 | Lower-cost edge systems | Less memory and compute capacity |
Functional safety is critical in factories, logistics, healthcare, and shared spaces. It differs from cybersecurity: it ensures predictable responses to hardware or software faults, reducing the risk of physical damage.
Nvidia has not yet disclosed final certifications, safety integrity levels, or detailed configurations for the IGX T3000. Its compatibility with Halos should be seen as part of the announcement, not as a universal certification for any robot built with the module.
Development can start with emulation
Physical modules will not be available until the first quarter of 2027, but Nvidia will allow early development through the existing Jetson AGX Thor kit.
The emulation mode for the T3000 will be released in July with JetPack 7.2.1. Emulation for the T2000 will follow later, with no specific date yet. This enables developers to test resource limits and approximate future module behavior.
Emulation helps verify whether an application fits in memory and achieves needed performance but does not replace testing on actual hardware. Factors like temperature, power, interfaces, latency, and sustained behavior may differ when the commercial modules arrive.
Nvidia already works with system integrators like ADLINK, Advantech, AAEON, Aetina, Seeed Studio, Connect Tech, and AVerMedia on Thor. It also mentions companies like 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot that are developing or deploying on the platform.
This launch marks a shift in the robotics market. Nvidia initially introduced Thor at the high end, with a platform featuring 128 GB and up to 2,070 teraflops FP4. The T3000 and T2000 lower specs aim at systems where cost, size, and power are more critical than maximum capacity.
Important data like pricing, exact power, dimensions, interface count, and performance on specific models are still awaited. Until technical sheets are published and independent tests are conducted, these new Jetson modules are best viewed as a promising expansion of the Thor family, not ready-for-deployment products.
FAQs
When will the Jetson T3000 and T2000 be available?
Nvidia plans to release both modules in Q1 2027. Prior to that, development can begin via emulation on the existing Jetson AGX Thor kit.
What’s the difference between T3000 and T5000?
The T3000 reduces memory from 128 GB to 32 GB and announced performance from 2,070 to 865 teraflops FP4. Nvidia also claims it is roughly half the size and power, though exact figures are not provided.
What is the purpose of the Jetson T2000?
It targets visual agents, mobile robots, manipulators, and edge systems that require Thor but not the memory or computational capability of higher-end modules.
What is Cosmos 3 Edge?
A 4-billion-parameter model designed to help robots and physical systems interpret their environment, reason, and generate actions locally using Nvidia Thor.
via: blogs.nvidia
