NVIDIA has just launched the Jetson T2000 and T3000 “mainstream” modules as smaller and lower cost options to the Jetson T4000/T5000 modules introduced last year for Edge AI and robotics applications.
The Jetson Thor T3000 and the industrial version IGX Thor T3000, which adds functional safety, both deliver up to 865 FP4 TFLOPS of AI compute. They feature a 1536-core NVIDIA Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory with 273GB/s bandwidth, and 25 GbE connectivity. The Jetson Thor T2000 delivers up to 400 TFLOPS of AI performance, features a 1024-core Blackwell GPU, 16GB LPDDR5 with 137GB/s bandwidth, and 10 GbE networking. Both are about half the size of the Jetson T4000/T5000 modules.
Here’s a preliminary comparison between the Jetson Thor T2000, T3000, T4000, and T5000 modules.
Jetson T2000 | Jetson T3000 | Jetson T4000 | Jetson T5000 |
|
|---|---|---|---|---|
| AI Performance | 400 TFLOPS (FP4—Sparse) | 865 TFLOPS (FP4—Sparse) | 1200 TFLOPS (FP4—Sparse) | 2070 TFLOPS (FP4—Sparse) |
| GPU | 1024-core NVIDIA Blackwell architecture GPU | 1536-core NVIDIA Blackwell architecture GPU | 1536-core NVIDIA Blackwell architecture GPU with 64 fifth-gen Tensor Cores Multi-Instance GPU (MIG) with 6 TPCs | 2560-core NVIDIA Blackwell architecture GPU with 96 fifth-gen Tensor Cores Multi-Instance GPU (MIG) with 10 TPCs |
| CPU | 6-core Arm Neoverse-V3AE 64-bit CPU | 8-core Arm Neoverse-V3AE 64-bit CPU 1 MB L2 cache per core | 12-core Arm Neoverse-V3AE 64-bit CPU 64 KB I-Cache, 64 KB D-Cache 1 MB L2 cache per core 16 MB Shared System L3 Cache | 14-core Arm Neoverse-V3AE 64-bit CPU 1 MB L2 cache per core 16 MB Shared System L3 Cache |
| Memory | 16 GB LPDDR5X 137 GB/s | 32 GB 256-bit LPDDR5X 273 GB/s | 64 GB 256-bit LPDDR5X 273 GB/s | 128 GB 256-bit LPDDR5X 273 GB/s |
| Networking | 2x 10GbE | "25GbE connectivity" | 3x 25GbE | 4x 25GbE |
| Mechanical | About 50 x 87 mm (half the size of the T4000/T5000 module | 100 x 87 mm; 699 pins |
||
| Power | Not specified | About half the power of T5000 (20 W–65 W?) | 40 W–75 W | 40 W–130 W |
It’s preliminary because, as I write this article, NVIDIA has yet to provide the full specifications for the new T2000 and T3000 modules. Instead, I relied on the announcement and the specifications for the upcoming AAEON BOXER-8752AI and BOXER-8723AI fanless embedded BOX PCs based on T2000 and T3000, respectively.
FLOPS numbers are one thing, but NVIDIA claims the T3000 achieves similar inference performance to the T5000 for multimodal workloads, including LLMs, VLMs, vision language action models, and world foundation models. So, depending on the application, migrating to T3000 can help reduce costs amid high memory prices without impacting performance in a significant way.
In any case, that means NVIDIA now has a scalable portfolio for Edge AI and robotics applications from the 70 TOPS Jetson Orin Nano up to the 2070 TOPS Jetson T5000.
There won’t be any specific Jetson Thor T2000/T3000 development kit, and developers can leverage the Jetson AGX Thor developer kit to emulate the performance of T3000 and T2000 modules. The T3000 will be supported in the Jetpack 7.2.1 SDK to be released later this month, and the T2000 emulation mode is expected in a future release. Both modules are scheduled to become available in Q1 2027.
Besides the hardware announcement, NVIDIA also released the Jetson agent skills, which developers can notably use to optimize the entire software stack and reduce memory usage in a matter of days. It’s not a small reduction either, as companies like UBTech, Agile Robots, and Connect Tech are said to have reduced memory usage by up to 15GB, enabling them to switch to the Jetson AGX Orin 32GB module, replacing the more expensive 32GB variant.
Another software release is the Cosmos 3 Edge 4-billion parameter model, designed to enable embodied systems to “see the world, reason over it in real time, and predict and generate actions through on-device inference”. It’s available through the NVIDIA Cosmos 3 open world foundation model family.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links.
