NVIDIA launches small mainstream Jetson T2000 and T3000 modules for edge AI and robotics applications

NVIDIA has launched Jetson T2000 and T3000 “mainstream” modules as a smaller, lower-cost option to the Jetson T4000/T5000 modules introduced last year for edge AI and robotics applications.

The Jetson Thor T3000 and the industrial version with added functional safety, the IGX Thor T3000, both offer up to 865 FP4 TFLOPS of AI computing. It features a 1536-core NVIDIA Blackwell GPU, an 8-core Neoverse Arm CPU, 32 GB of LPDDR5X memory with 273 GB/s bandwidth, and 25 GbE connectivity. Jetson Thor T2000 delivers up to 400 TFLOPS of AI performance and features a 1024-core Blackwell GPU, 16 GB LPDDR5 with 137 GB/s bandwidth, and 10 GbE networking. Both are about half the size of Jetson T4000/T5000 modules.

Here is a preliminary comparison between 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 5th Generation Tensor Cores Multi-instance GPU (MIG) with 6 TPCs	2560-core NVIDIA Blackwell Architecture GPU with 96 5th Generation 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 1MB 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	16GB LPDDR5X 137GB/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 connection”	3x 25GbE	4x 25GbE
mechanical	Approximately 50 x 87 mm (half the size of T4000/T5000 modules)		100×87mm; 699 pins
force	not specified	About half the power of T5000 (20W to 65W?)	40W~75W	40W～130W

As of this writing, NVIDIA has not yet provided full specifications for the new T2000 and T3000 modules, so this is preliminary. Instead, I relied on the announcement and specs of the upcoming AAEON BOXER-8752AI and BOXER-8723AI fanless embedded BOX PCs, based on the T2000 and T3000, respectively.

While FLOPS numbers are important, NVIDIA claims the T3000 achieves similar inference performance to the T5000 for multimodal workloads such as LLM, VLM, Vision Language Action Model, and World Foundation Model. Therefore, for some applications, migrating to the T3000 can reduce costs even as memory prices continue to rise without significantly impacting performance.

Either way, this means NVIDIA has a scalable portfolio for edge AI and robotics applications, from the 70 TOPS Jetson Orin Nano to the 2070 TOPS Jetson T5000.

There is no specific Jetson Thor T2000/T3000 development kit, and developers can utilize the Jetson AGX Thor development kit to emulate the performance of T3000 and T2000 modules. T3000 will be supported in the Jetpack 7.2.1 SDK released later this month, and T2000 emulation mode will be supported in a future release. Both modules are expected to be available in Q1 2027.

In addition to the hardware announcement, NVIDIA also released Jetson agent skills. Developers can specifically use it to optimize their entire software stack and reduce memory usage in days. This is no small reduction, and companies like UBTech, Agile Robots, and Connect Tech are said to have reduced memory usage by up to 15GB, allowing them to switch from more expensive 32GB versions to Jetson AGX Orin 32GB modules.

Another software release is the Cosmos 3 Edge 4-billion-parameter model, which is designed to enable materialized systems to “perceive the world, reason in real-time, and predict and generate actions through on-device inference.” This is available through the NVIDIA Cosmos 3 open world foundation model family.

Support CNX Software! Donate in crypto, become a patron on Patreon, or buy on Amazon or Aliexpress. We also use affiliate links in our articles and receive a commission if you click the link and purchase a product.