NVIDIA has launched the Jetson T2000 and T3000 modules as compact, mainstream options that deliver high AI computing power for edge and robotics use cases while reducing size and cost compared to last year’s T4000 and T5000 models.
The Jetson Thor T3000 module, along with its industrial variant IGX Thor T3000 that adds functional safety features, supports up to 865 FP4 TFLOPS of AI calculation power. It integrates a 1536-core NVIDIA Blackwell GPU, an eight-core Arm Neoverse CPU, 32GB of LPDDR5X memory with bandwidth reaching 273GB/s, and 25 GbE network connectivity. Meanwhile, the Jetson Thor T2000 delivers up to 400 FP4 TFLOPS, equipped with a 1024-core Blackwell GPU, 16GB LPDDR5 memory at 137GB/s, and dual 10 GbE ports.
Both modules measure approximately half the size of previous T4000 and T5000 units, making them suitable for space-constrained applications. Power consumption is expected to be roughly half that of the T5000, though official figures are yet to be confirmed.
NVIDIA positions the T3000 as an efficient alternative for workloads involving multimodal AI models—such as large language models (LLMs), vision-language models (VLMs), and world foundation models—offering similar inference performance as the larger T5000 while reducing hardware expenses, especially amid rising memory costs.
The company’s Edge AI portfolio now spans a wide range of capabilities, from the entry-level Jetson Orin Nano with 70 TOPS, through the new T2000 and T3000 modules, up to the high-end T5000 with 2070 TOPS. Developers won’t have dedicated development kits for the T2000 or T3000; however, they can simulate these modules’ performance using the existing Jetson AGX Thor developer kit.
Networking capabilities differ across models, with the T2000 featuring dual 10 GbE ports, while T3000 supports a single 25 GbE connection. Memory capacity and CPU cores also increase from the T2000 to the T3000, reflecting their targeted application ranges.