DGX Spark vs Jetson AGX Thor
We rent Sparks, and this page still has to say the honest thing: the Spark and the Jetson AGX Thor are not really cross-shopped by the same buyer. Both are compact NVIDIA Arm boxes built on recent GPU architectures, and that surface similarity is where the overlap mostly ends. We are not confident of Thor's exact current memory, bandwidth and compute specifications, so rather than print numbers we cannot stand behind, this page stays qualitative on that front and says plainly where we are hedging.
Two different target markets
The DGX Spark is built and sold as a desk-sized LLM development and inference machine: DGX OS ships a full Ubuntu desktop, it plugs into a monitor and keyboard, and NVIDIA positions it explicitly as a way to prototype on the same CUDA stack that runs in a datacenter, detailed on GPUwerk's hardware page.
The Jetson AGX Thor sits in NVIDIA's Jetson line, which has always targeted embedded and edge deployment: robotics, autonomous machines, industrial vision systems and similar workloads where the compute board is a component inside a larger physical product, not a workstation someone sits at. It is designed to be embedded and run headless as part of a robot or vehicle's control system, a genuinely different deployment shape from a desk box with a display output.
That difference in target market matters more than any spec-sheet number: buying a Thor to do what a Spark does, or the reverse, means fighting the platform's intended deployment shape.
Specs: what we are and are not confident about
We are not highly confident of Jetson AGX Thor's exact current memory capacity, memory bandwidth or peak compute figures as of this page's publish date, and print specifications change across a product's revisions and configurations. Rather than state numbers we have not verified against NVIDIA's current documentation, this comparison stays qualitative on those axes. If you need exact current Thor specifications, check NVIDIA's own Jetson AGX Thor product and developer pages directly rather than trusting a number on this page.
What we are confident stating, because it follows from each platform's stated purpose rather than a spec we would need to verify:
- Form factor. The Spark is a standalone desktop unit with display output, meant to sit on a desk and be used directly. Thor is a compute module or carrier-board system meant to be integrated into another product's chassis, not used as a standalone desktop.
- Software stack emphasis. The Spark's DGX OS is a general-purpose Ubuntu desktop with the full CUDA, PyTorch, vLLM, TensorRT-LLM and NeMo stack aimed at LLM development, per GPUwerk's hardware page. Thor's software emphasis is NVIDIA's robotics and autonomy tooling, built around perception, planning and control workloads rather than LLM serving, though it shares the same underlying CUDA foundation as an NVIDIA Arm platform.
- Networking for LLM clustering. The Spark ships two ConnectX-7 QSFP ports specifically for linking two units into a 256 GB pool for larger LLM workloads, per GPUwerk's hardware page. GPUwerk is not aware of an equivalent LLM-clustering feature being the design intent for Thor, whose networking and I/O are oriented toward sensor and actuator integration in an embedded system instead.
If your workload is actually LLM development or serving
The Spark is the right box to evaluate, and it is the one GPUwerk has measured directly: single-user and concurrency tok/s figures across six models are on the benchmarks page, and the arithmetic for predicting a new model's speed before downloading it is there too. Nothing on this page changes that recommendation; it exists to say plainly that Thor is not a like-for-like alternative for this job, not to talk you out of the Spark.
If your workload is robotics or embedded edge inference
That is Thor's actual design center, and GPUwerk does not sell, rent or have measured experience with it: we operate a Spark fleet for LLM development and inference, not an embedded robotics practice. If that is your workload, NVIDIA's own Jetson documentation and developer resources are the right starting point rather than anything on this comparison page.
Unsure which category your project actually falls into? A first engagement can help think through whether an LLM-serving box like the Spark is even the right category of hardware for what you're building.