NVIDIA this week announced it will expand its Jetson Thor lineup of robotics and edge AI boards with a pair of new mid-grade models. Coming to the Jetson lineup at the start of 2027 will be the Jetson Thor T3000 and T2000, which will join the existing Jetson Thor boards as lower-cost, lower-performance, lower-power alternatives to NVIDIA’s current high-end boards. This hits particularly close to home since our NVIDIA Jetson AGX Thor Developer Kit runs an amazing array of applications 24×7 in the STH studio.
Newer, Cheaper Jetson Thor Boards: T3000 and T2000 #
NVIDIA’s latest Jetson Thor announcement comes as, although the company is greatly benefiting from the boom in AI, the spike in component costs is hurting their other business, both in the consumer and in the industrial space. For the Jetson Thor boards, the only options for customers specifically after NVIDIA’s Blackwell-based Thor hardware are to buy the existing T4000 and T5000 boards, which ship with 64GB and 128GB of memory, respectively. To that end, the company is taking a couple of important steps to expand the Jetson Thor hardware lineup while addressing the cost concerns brought on by memory prices.
First and foremost, then are the new Jetson T3000 and T2000 boards. Both of these Jetson boards are based on the same Thor SoC that powers the rest of the current Thor products, but are further pared down in terms of configuration and capabilities.
NVIDIA Jetson Thor Lineup | |||| T2000 | T3000 | T4000 | T5000 | | CPU Cores (Neoverse-V3AE) | 6 | 8 | 12 | 14 | CUDA Cores | 1024 | 1536 | 1536 | 2048 | GPU Perf(FP4 Sparse) | 400 TFLOPS | 865 TFLOPS | 1200 TFLOPS | 2070 TFLOPS | Memory(LPDDR5X-8500) | 16GB | 32GB | 64GB | 128GB | Memory Bandwidth | 137GB/sec | 237GB/sec | 237GB/sec | 237GB/sec | Networking | 10GbE | 25GbE | 3x 25GbE | 4x 25GbE | Max TDP | ? | ~65W | 70W | 130W | Release Date | Q1’2027 | Q1’2027 | Q1’2026 | Q3’2025 |
The more powerful of the two new boards will be the T3000, which offers 8 Arm Neoverse V3AE CPU cores along with a 1536-core Blackwell iGPU. In terms of performance, NVIDIA is quoting 865 TFLOPS of spare FP4, which is about 72% of the T4000’s GPU performance. Notably, both parts feature identical GPU configurations, so the performance difference would come down to GPU clock speeds rather than reducing the amount of GPU hardware further.
The Thor SoC on the T3000 board will be paired with 32GB of memory, half the amount found on the T4000. Importantly, despite the reduction in memory capacity, this board retains the full memory bandwidth of the Thor SoC: LPDDR5X-8500 on a 256-bit memory bus, providing 273GB/s of memory bandwidth. While the T3000 has lower GPU performance than the more powerful Jetson boards, it offers the same memory bandwidth they provide.
It is presumably for this reason that NVIDIA is promoting the T3000 as offering “similar inference performance of the T5000 for multimodal workloads.” For AI workloads that are primarily memory bandwidth-bound (not memory *capacity-*bound), the T3000 is likely to be quite close to the T4000 and T5000 in that regard. This, in turn, is why NVIDIA is pitching the T3000 as a more budget-friendly option for customers who need Thor-based boards for their products but are held back by the high (and rising) costs of current high-end Jetson boards.
Rounding out the package, the T3000 retains Thor’s 25Gb Ethernet connectivity, though it is not clear how many of the SoC’s native 4 controllers will be exposed. The complete board is said to consume about half the power of the T5000, or 65 Watts.
Overall, NVIDIA is aiming the T3000 at the robotics sector as a more cost-efficient option for those markets. Meanwhile, the company will also be offering an IGX version of the board for customers who need a T3000 with functional safety support as well.
Joining the T3000 is the T2000, a further cut-down version of the board, priced lower. The T2000 will be a 6-core CPU and 1024 CUDA core configuration, rated to deliver 400 TFLOPS of sparse FP4, or about 46% of the performance of the more powerful Jetson board. The Thor SoC in this configuration will be paired with just 16GB of LPDDR5X running on half of the memory bus, giving it a total memory bandwidth of 137GB/second.
Meanwhile, this board is also taking a trim in networking capabilities. Instead of 25GbE networking, it will offer 10GbE networking.
Compared to the T3000, the T2000 is significantly less powerful in many ways, making it a more distinctive product. Fittingly, then, NVIDIA says it will primarily pitch the T2000 at lighter-weight edge AI applications rather than complex robotics applications.
Memory-Optimized Jetson Skills #
As well as lower-cost Jetson Thor configurations, NVIDIA is also addressing the impact of high memory prices on its Jetson hardware lineup by addressing memory pressure on the software side. Specifically, the company is releasing a new set of highly optimized agent skills that will require less memory than before.
In its press release, NVIDIA is surprisingly blunt on the matter, stating that they are doing this to enable its customers to be able to move down to a lower memory configuration within the same hardware tier while still being able to run their workloads at similar performance.
NVIDIA offered several examples of customers who reduced their memory consumption enough to move to a smaller memory configuration for robotics workloads on Orin platforms. Or, in one case, saving enough memory on the several-year-old Jetson TX2 in order to add further functionality to the product.
Final Words #
Once they launch next year, the new Jetson Thor boards will be joining the rest of NVIDIA’s ongoing Thor lineup. With the new T2000 and T3000 boards bringing down the Thor hardware to a lower price than the $3000 T4000, NVIDIA will retain its previous-generation Orin hardware to flesh out the lower half of the Jetson lineup. This includes the AGX Orin and Orin NX for the rest of the mainstream products, along with the Orin Nano as their entry-level product.
NVIDIA plans to release the Jetson T3000 and T2000 boards in Q1 of 2027. Pricing is not being announced at this time.