NVIDIA now sells two things with "Spark" in the name, and they are easy to mix up. The DGX Spark is a small desktop box for AI work. The RTX Spark is a chip for Windows laptops and compact desktops. They share a family and a memory design, but they do different jobs. Which one fits depends on where you work and what you run.
What they share
Both pair an Arm CPU (NVIDIA calls it Grace) with a Blackwell GPU, and both use one pool of unified LPDDR5X memory that the CPU and GPU share. That is what lets a small machine hold a model that would not fit on a consumer graphics card with 32 GB. The DGX Spark uses the GB10 chip. The RTX Spark uses a chip NVIDIA calls N1X.
What NVIDIA publishes
| Item | RTX Spark (N1X) | DGX Spark (GB10) |
|---|---|---|
| CPU, GPU | 20-core Grace with 6,144-core GPU, or 18-core with 5,120-core GPU | 20-core Arm with Blackwell GPU |
| Memory | Up to 128 GB (20-core chip), up to 64 GB (18-core chip) | 128 GB, or 64 GB through partners |
| Memory bandwidth | Not published | 273 GB/s |
| Operating system | Windows 11 | NVIDIA DGX OS (Ubuntu-based Linux) |
| Form | Slim laptops, compact desktops | Desktop box |
The RTX Spark bandwidth is the important gap. Without it, speed cannot be estimated, and the EMD Lab selector shows no speed for it. We would rather say that plainly than guess. Writing speed for chat, tok/s (tokens per second, roughly words per second), follows memory bandwidth, so once NVIDIA or a maker publishes it, the estimate follows.
What that means in practice
- Laptops run the chip at lower power. In a thin laptop, the chip has less power and cooling to work with than in a desktop box. Expect long jobs that need a lot of computing, such as reading long documents or fine-tuning, to be slower in a laptop. Short chat answers depend more on memory and may hold up better. No reliable head-to-head public test exists yet.
- Memory size decides what runs. The selector puts gpt-oss-120b at about 61 to 67 GB with its context. A 64 GB machine cannot hold it. Only the 128 GB RTX Spark configurations can, so check the memory before you check anything else.
- Serving many users. vLLM, the software that serves many users at once, runs inside WSL (Windows Subsystem for Linux) on Windows. On the DGX Spark it runs on DGX OS directly.
- No linking. The DGX Spark carries a ConnectX-7 network port rated at 200 Gb/s, so two units can be linked to share a larger model. The RTX Spark page lists no such port.
- Windows on Arm is young. Many Windows apps run, but some tools, drivers and games are still catching up on Arm. Check the tools you need.
The real-world anchor for the DGX side
For the DGX Spark, the llama.cpp community has logged gpt-oss-120b runs with CUDA. These are public results, not EMD Lab measurements.
| Build | Reads a prompt | Writes | Writes, 32K deep |
|---|---|---|---|
| 22 Oct 2025 | 1,737 | 45.9 | 34.3 |
| 2 Nov 2025 | 1,956 | 60.6 | 40.6 |
"Reads" is prompt processing, how fast the box digests your input. "32K deep" means 32,000 tokens are already in the context. The selector estimates about 42 tok/s for gpt-oss-120b on the GX10 at 32K, in the same range as those results. For the RTX Spark we have no equivalent public data to quote.
Who each one is for
RTX Spark: a person who wants one Windows laptop that also has room for big local models. It suits someone on the move who also wants to try a 100B-class model on their own machine, and who is happy with Windows on Arm. The memory choice matters more than anything else: 128 GB for the 120B class, less for smaller models.
DGX Spark and the ASUS Ascent GX10: developers who work in CUDA, and small teams that want a shared experiment machine on a desk. It runs DGX OS, has the 200 Gb/s port for linking two units, and does not depend on a laptop's power limits. The GX10 is ASUS's GB10 box; we carry ASUS, and you can ask us about it through the quote form. We do not list it as an item on the site.
If your team is more than a handful of people, or you need fast answers at long context, neither is the right tool; look at a GPU server such as the QNAP AI server. Our hardware guide compares the platforms, and the model finder shows what fits.
The short version
- Same family: Grace Arm CPU, Blackwell GPU, unified LPDDR5X. RTX Spark runs Windows 11; DGX Spark runs DGX OS.
- RTX Spark memory bandwidth is not published, so its speed cannot be estimated yet. DGX Spark is 273 GB/s.
- Memory size decides what runs: a 64 GB machine cannot hold gpt-oss-120b.
- Laptops run at lower power, so long jobs will be slower; there is no 200 Gb/s port; vLLM needs WSL.
- RTX Spark suits the Windows laptop user; DGX Spark and the GX10 suit CUDA developers and small-team desks.