NVIDIA Jetson AGX Thor T4000 64GB NEW

The T4000 is the smaller of the two Jetson Thor modules, and it is a clean three-fifths of the T5000 rather than a different design: 1,536 Blackwell CUDA cores against 2,560, 12 Arm Neoverse-V3AE cores against 14, and 64 GB of LPDDR5X against 128 GB. The interesting part is what it does not give up. Memory stays on the same 256-bit bus at the same 273 GB/s, so the smaller module has more bandwidth per core than the larger one, and the clock is actually higher in its default mode, 1.53 GHz against the T5000's 1.386 GHz at 120 W. It delivers 600 dense FP4 TFLOPS and 300 dense FP8 within a 90 W ceiling, where the T5000 needs 130 W for its 1,035 and 517. For a robot or an edge box that has to run a real vision language action model on a battery or a fixed power budget, the T4000 is often the module that fits, and 64 GB is still enough memory for a model that would not go near an Orin. MIG is supported, so one module can be partitioned across several workloads.