RTX Spark and DGX Spark are two different NVIDIA products with similar names. DGX Spark is a finished computer: a small Linux desktop that NVIDIA has sold since October 2025, now priced at $4,699 with 128GB of unified memory. RTX Spark is a chip platform: NVIDIA supplies it to PC makers, who build Windows laptops and compact desktops around it, with first shipments in October 2026 and no announced price. The top RTX Spark configuration matches DGX Spark on core counts and memory. The differences are the operating system, the form factor, who sells it, and how much has been measured.
Side by Side
| RTX Spark | DGX Spark | |
|---|---|---|
| What it is | A chip used in PCs from ASUS, Dell, HP, Lenovo, Microsoft, MSI, and others | A mini desktop sold by NVIDIA and partners |
| Operating system | Windows 11 on Arm | NVIDIA DGX OS, based on Ubuntu Linux |
| CPU | 20-core Grace, or 18-core in the second configuration | 20-core Arm: 10 Cortex-X925 and 10 Cortex-A725 |
| GPU | Blackwell RTX, 6,144 or 5,120 CUDA cores | Blackwell, 6,144 CUDA cores |
| Unified memory | 24GB to 128GB | 128GB |
| Memory bandwidth | About 300 GB/s reported, not in NVIDIA's announcement | 273 GB/s, NVIDIA specification |
| Storage and networking | Set by each PC maker | 4TB NVMe, 10 GbE, and a 200 Gbps ConnectX-7 port |
| Size | 14 to 16 inch laptops from 14 mm thick, plus compact desktops | 150 by 150 by 50.5 mm, 1.2 kg |
| Power | Not published | 240 watt supply, 140 watt TDP |
| NVIDIA model-size claim | 120-billion-parameter models, 1 million token context | Inference up to 200 billion parameters, fine-tuning up to 70 billion |
| Price | Not announced. Analyst estimate about $2,899 for a high-end system | $4,699, up from $3,999 at launch |
| Availability | First systems ship October 2026 | On sale since October 2025 |
| Independent LLM benchmarks | None as of October 1, 2026 | Many |
RTX Spark figures come from NVIDIA's announcement and from Wccftech's IFA report. DGX Spark figures come from NVIDIA's product page, and the CUDA core count and price history from IntuitionLabs' review.
Measured Speed Exists Only for DGX Spark
Every tokens-per-second figure in circulation for a "Spark" is a DGX Spark figure. Three testers have published single-stream numbers.
| Model and format | Prompt processing (tokens/sec) | Generation (tokens/sec) | Measured by |
|---|---|---|---|
| gpt-oss 120B, MXFP4, empty context | 1,956 | 60.6 | llama.cpp scoreboard |
| gpt-oss 120B, MXFP4, 32K context | 1,027 | 40.6 | llama.cpp scoreboard |
| gpt-oss 20B, MXFP4, empty context | 2,009 | 60.9 | llama.cpp scoreboard |
| gpt-oss 20B, MXFP4 | 2,053 | 49.7 | LMSYS, with Ollama |
| Gemma 3 27B, Q4_K_M | 834 | 10.8 | Ollama |
| Llama 3.1 8B, Q4_K_M | 7,614 | 38.0 | Ollama |
| Llama 3.1 70B, Q4_K_M | 1,911 | 4.4 | Ollama |
The spread is wide. Sparse models in NVIDIA's preferred 4-bit format run at 40 to 61 tokens per second. Dense models in standard GGUF quantisation are far slower, down to 4.4 tokens per second at 70B. For scale, LMSYS measured an RTX 5090 at 205 tokens per second on the same gpt-oss 20B test where DGX Spark reached 49.7.
These numbers are from October 2025. NVIDIA has shipped software updates since, and a February 2026 llama.cpp run cited by AIMultiple shows gpt-oss 20B at 83.4 tokens per second at empty context. See the cross-device benchmark table for the wider context.
Will RTX Spark Match DGX Spark
Nobody has tested it. The top RTX Spark configuration has the same core counts and memory ceiling, which suggests similar speed when cooling allows. Three things could move the result:
- Cooling. DGX Spark is a 1.2 kg box with a 140 watt TDP. Leaked documents put RTX Spark laptops at 45 to 80 watts, which NVIDIA has not confirmed. Microsoft's Dev Box is described with a 100 watt thermal envelope.
- Software. DGX Spark benchmarks run on Linux. RTX Spark runs Windows on Arm, and Microsoft says its Dev Box supports CUDA through WSL 2 with GPU passthrough.
- Memory tier. RTX Spark starts at 24GB. Only the 128GB tier is comparable to DGX Spark for large models.
Which One Fits Which Buyer
| Need | Better fit | Reason |
|---|---|---|
| A machine to buy today | DGX Spark | Shipping, priced, and benchmarked |
| A laptop | RTX Spark | DGX Spark is desktop only |
| Windows applications and games | RTX Spark | DGX Spark runs Linux |
| Linking two units | DGX Spark | It has a 200 Gbps ConnectX-7 port |
| Fine-tuning on the standard Linux stack | DGX Spark | NVIDIA lists fine-tuning up to 70 billion parameters |
| Lowest price for 128GB | Unknown | RTX Spark prices are not announced |
Buyers comparing either Spark with Apple hardware should read the Mac Studio M5 Ultra page, which covers DGX Spark vs M5 Ultra, and the older DGX Spark vs M5 Max comparison. The single-device view is on the RTX Spark page.
What Is Still Unknown
RTX Spark prices, official memory bandwidth, sustained power in a laptop, and any tokens-per-second result are all unpublished. It is also not confirmed whether NVIDIA's NVFP4 and TensorRT optimizations for DGX Spark will ship in the same form on Windows.
Brand Visibility Implications
DGX Spark put 128GB of model memory on developers' desks. RTX Spark aims to put it in ordinary Windows laptops. Either way the model answers from its weights, with no retrieval unless the user adds it. What the model learned about a brand in training is what the user hears, and no analytics tool sees the question. See the local LLM visibility blind spot.
Methodology
This page compiles vendor specifications and third-party benchmark reports. None of the figures are Presenc AI measurements. Vendor facts come from primary pages, including NVIDIA's RTX Spark announcement and DGX Spark product page, and Apple's Mac Studio announcement. Throughput figures come from named testers: the llama.cpp scoreboards for CUDA cards, DGX Spark and Apple Silicon, Hardware Corner, MacStories, LMSYS, the Ollama blog, StorageReview, and the Level1Techs forum. Each number is attributed in the text to whoever measured or claimed it. Testers use different models, quantisation, runtimes, and context lengths, so compare figures within one source and treat cross-source comparisons as approximate. Prices are as reported in late September 2026 and are moving with memory supply. Status as of October 1, 2026.
How Presenc AI Helps
Presenc AI tracks how AI models describe and recommend your brand, including the open-weight model families that run on machines like these. That gives a view of answers that never touch a cloud endpoint.