Use NVIDIA’s first-boot wizard, then start with one local inference runtime before adding notebooks, applications or multi-device jobs. The machine commonly called “RTX Spark” is documented by NVIDIA as DGX Spark: a compact Grace Blackwell system with a 20-core ARM64 processor, integrated GPU and 128 GB of unified memory. It ships with DGX OS, CUDA, cuDNN, Docker and NVIDIA Container Runtime, so the main setup work is networking, updates, remote access and choosing a model runtime.
This guide covers the direct-display and network-appliance setup paths, the runtime choice between llama.cpp, Ollama and LM Studio Bionic, Windows and Windows Subsystem for Linux considerations, and a safe progression from one-device inference to larger workflows.
As an Amazon Associate I earn from qualifying purchases.
What NVIDIA RTX Spark is and why it matters for local AI
NVIDIA’s official product name is DGX Spark. “RTX Spark” is a common shorthand, but the setup screens, operating-system documentation and developer recipes use DGX Spark. NVIDIA describes it as a compact AI computer for developers, data scientists and AI researchers who need local AI development and deployment capability.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsThe system combines a Grace Blackwell architecture with a 20-core ARM64-based processor and an integrated GPU. Its defining resource is 128 GB of unified memory: CPU and GPU workloads draw from the same pool rather than from separate system RAM and graphics memory. That design is intended to make substantially larger local models practical than on an ordinary desktop GPU, but it does not guarantee a particular generation speed or context length.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
- Memory: 128 GB unified memory.
- CPU: 20-core ARM64-based processor.
- Connectivity: Wi-Fi 7 and 10 GbE, with ConnectX-7 networking.
- Physical size: 150 mm × 150 mm × 50.5 mm desktop enclosure.
- Software included: DGX OS, CUDA, cuDNN, Docker and NVIDIA Container Runtime.
NVIDIA documents workloads reaching up to 200 billion parameters and multi-device scaling. Read those statements as platform capabilities, not as a promise that every 200-billion-parameter model will fit at a useful speed. Quantization, model architecture, context size, runtime, thermals and the workload all change the result.
Prepare the hardware before first power-on
You can configure Spark as a small workstation with its own display, keyboard and mouse, or treat it as a network appliance and complete setup from another computer.
Have these items ready
- DGX Spark and its power connection.
- Network access, using Ethernet or Wi-Fi.
- For local setup: a display, keyboard and mouse.
- For network-appliance setup: another computer or mobile device capable of joining a temporary Wi-Fi network and opening the setup wizard.
- The supplied Quick Start Guide, because the temporary hotspot SSID and password are printed there.
Connect the system to power and have the network available before starting. If you choose the hotspot route, keep the Quick Start Guide nearby; the temporary network credentials are not something to infer or substitute.
Complete the first-boot wizard
NVIDIA documents two equivalent setup routes. Select the one that matches how you intend to use Spark.
Route A: local display, keyboard and mouse
- Connect a display, keyboard and mouse to Spark.
- Power on the unit.
- Follow the on-screen wizard to select your language, time zone and keyboard layout.
- Allow the system to install updates and join your home or office network.
- Wait for the initial update and network process to finish. NVIDIA says this stage can take as long as 10 minutes.
- When the wizard completes, sign in locally and open the DGX Dashboard to confirm that the system is visible and ready.
This route is the simplest recovery path because you retain a direct console if networking or remote access later fails.
Route B: temporary Wi-Fi hotspot
- Power on Spark without connecting a permanent display setup.
- On another computer, join the temporary Wi-Fi hotspot whose SSID and password are printed in the supplied Quick Start Guide.
- Open the browser-based setup wizard presented by Spark.
- Set the language, time zone and keyboard settings.
- Use the wizard to connect Spark to the intended home or office network.
- Leave the unit powered while updates install and the network transition completes; NVIDIA allows up to 10 minutes for this stage.
- After Spark joins the permanent network, disconnect the other computer from the temporary hotspot and use the address or discovery method shown by the setup interface for normal access.
The hotspot method is convenient when Spark will live near your router or in a lab without a dedicated monitor. It depends on the temporary wireless network during setup, so do not switch away from it until the wizard confirms the permanent connection.
Orient yourself after setup
Once first boot is complete, you can use the machine locally or remotely. NVIDIA lists local access, SSH, NVIDIA Sync and remote desktop as available post-setup approaches. Keep at least one working access method before removing the display and keyboard.
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
Use DGX Dashboard for administration
DGX Dashboard is the central starting point for monitoring, updates and access to JupyterLab. Use it to check that the system is healthy before installing a model runtime. It is also the place to look when an update, service or notebook launch behaves differently from the initial setup.
Keep a local recovery path
Remote access is more convenient, but a directly attached display and keyboard are valuable if the network address changes, an SSH service is unavailable or an update interrupts normal access. If Spark is deployed as an appliance, record its network identity and retain a way to reconnect locally during maintenance.
Understand the software baseline and architecture
The current porting guidance describes a Ubuntu 24.04-derived base with a Linux 6.11-based NVIDIA Base OS. The preinstalled NVIDIA stack is designed for CUDA workloads rather than for a generic desktop installation.
For development from another machine, NVIDIA’s RTX Spark guidance also supports Windows and Windows Subsystem for Linux. These environments use the same broad classes of frameworks used in cloud deployments, allowing an application or model workflow to be developed on a Windows workstation and then moved to Spark. The processor architecture remains important: Spark itself is ARM64, so choose ARM64-compatible packages, containers and binaries whenever a project offers separate builds.
Software families available for the platform
NVIDIA’s software map includes PyTorch, llama.cpp, Ollama, Unsloth, vLLM and CUDA-X libraries. The DGX Spark developer portal adds recipes for ComfyUI image and video workflows, SGLang, NIM containers, fine-tuning, NCCL and connections between multiple Spark systems.
Do not install every framework on the first day. A single inference runtime gives you a known-good baseline. Add a notebook, application server or fine-tuning stack only after that baseline can load a model and answer a test prompt.
Choose a first local AI runtime
The practical choice is between control and convenience. NVIDIA currently recommends llama.cpp when you want direct control over speed, memory use and hardware behavior. Ollama is the simpler command-oriented option, while LM Studio Bionic offers a graphical download-and-test workflow.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
| Runtime | Best starting use | What the guidance establishes | Trade-off |
|---|---|---|---|
| llama.cpp | Controlled, inspectable inference | NVIDIA calls it the recommended backend when the developer wants control over speed, memory and hardware. The current recipe recommends Qwen3.8-27B for systems in the 24–128 GB range. | More setup decisions and more visible hardware or quantization details. |
| Ollama | Quick command-oriented model use | Listed by NVIDIA as a simpler option for running models through commands. | Less direct control than a hand-tuned llama.cpp workflow. |
| LM Studio Bionic | Graphical model discovery and testing | Provides a graphical way to download and test models. | Convenience can hide some of the runtime and memory decisions you may later need to tune. |
All three should be treated as alternative first runtimes, not layers that must be installed together. Running several servers at once can consume unified memory and make it harder to identify which process is responsible for a failure.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchStart with llama.cpp and a known model target
For a first controlled workload, use the current NVIDIA llama.cpp recipe and its stated Qwen3.8-27B recommendation for 24–128 GB systems. The exact command and quantization choice are maintained in that recipe and should be copied from its current version rather than from an old article or shell history.
- Open the current NVIDIA Build llama.cpp recipe for RTX Spark.
- Choose the build that matches your operating system and architecture. On Spark itself, that means an ARM64-compatible Linux path.
- Follow the recipe’s current installation and model-download instructions.
- Select the quantization specified for your intended memory and context target. Do not assume that a model’s parameter count alone determines fit.
- Launch the recipe’s server or command-line client and send a short test prompt.
- Record whether the model loads, the available context setting and the observed responsiveness before changing any other variable.
The point of this first run is not to claim a benchmark. It establishes that the operating system, CUDA stack, runtime and model can work together. Once that is stable, you can change quantization, context or batching one variable at a time.
Windows ARM64 requirement
If you are using the Windows ARM64 route described by NVIDIA, use the CUDA 13 build and place the CUDA 13.4 DLLs in the same folder as both llama-server.exe and llama-cli.exe. The placement requirement is specific to that route; copying binaries into a different directory layout can produce missing-DLL errors even when the model files are correct.
When Ollama or LM Studio is the better first step
Choose Ollama for a short command path
Ollama is appropriate when the goal is to get a local model endpoint running with minimal command-line ceremony. It is a good fit for an application developer who wants to test prompts or connect a small prototype before learning the lower-level controls exposed by llama.cpp.
Keep the first Ollama test isolated from other inference servers. If the model does not load, stop there and inspect the runtime’s own error output before adding Docker containers, notebook kernels or a second model.
Choose LM Studio Bionic for a graphical workflow
LM Studio Bionic suits users who prefer to browse, download and test models in a graphical interface. It can shorten the first experiment for someone who does not yet have a preferred shell workflow. After a model works, document its model choice, quantization and context settings so the result can be reproduced outside the graphical session.
Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
Connect an application, notebook or coding agent
After one model responds reliably, turn the runtime into a development building block. Start with one consumer: an application, a JupyterLab notebook or a coding agent. Verify the endpoint from that consumer before adding another.
- Confirm the inference server remains reachable after a fresh login or reboot.
- Record the endpoint address and the runtime’s API mode in your project configuration.
- Send a small deterministic test prompt from the application or notebook.
- Check response completion, context handling and error reporting.
- Only then increase context, add retrieval data, or move to a larger model.
DGX Dashboard provides access to JupyterLab, making notebooks a natural next step after the first model test. The same single-model endpoint can support a local prototype, but keep credentials and network exposure appropriate to your environment before allowing access beyond your trusted network.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Use Spark remotely without losing recovery options
Remote work is one of Spark’s strengths, but each access method has a different failure mode.
| Access method | Initial requirement | Convenience | Recovery consideration |
|---|---|---|---|
| Direct local access | Display, keyboard and mouse | Immediate console access without depending on another machine or network address. | Requires peripherals at the Spark location. |
| SSH | Spark must be reachable on the network and the SSH service must be available. | Efficient for shell work, scripts and repeatable development. | Keep local console access available for network or service failures. |
| NVIDIA Sync | A configured network connection and supported client workflow. | Designed for remote interaction within NVIDIA’s Spark tooling. | Behavior and setup can change with platform releases. |
| Remote desktop | A configured desktop session and network path. | Useful when a graphical application such as LM Studio or a desktop notebook is required. | More sensitive to session, display and network conditions than a shell connection. |
For a reliable pattern, complete first boot locally or through the temporary hotspot, confirm the permanent network, test SSH or another remote method, and only then move Spark to its final location. If a remote route fails, return to the direct console or the setup method documented for the current release.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Move from one device to larger workloads
Stay with one-device inference until you can reproduce a successful model load and application request. NVIDIA’s developer portal includes recipes for fine-tuning, vLLM, SGLang, NIM containers, NCCL and multi-Spark connections, but those components add operational variables.
When to add a notebook or fine-tuning tools
Add PyTorch, Unsloth or a notebook workflow when your objective changes from inference to experimentation or adaptation. Keep the original inference test as a control so you can tell whether a new dependency changed the base environment.
Free tools Windows power users keep installed
One-click scans. No signup required.
When to consider multi-device scaling
Move to multiple Spark systems only when a single system is stable and the model or workload genuinely needs more capacity. NCCL and multi-Spark recipes address communication between devices; they do not turn an unverified single-node setup into a predictable cluster. Validate network configuration, software versions and failure recovery before scheduling long jobs.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Capacity, speed and memory decisions
The 128 GB unified-memory design is the reason Spark can target large local models. A model’s parameter count is still only the beginning of the calculation. Weights, quantization metadata, runtime buffers, key-value cache and context length all consume memory. A larger context or concurrent requests can reduce the headroom available for model loading.
- Use the model recommendation in NVIDIA’s current recipe as a starting point, not as a universal performance claim.
- Change one variable at a time: model, quantization, context, batch size or concurrency.
- Watch memory and thermal behavior while the workload is active.
- Do not compare token-per-second results unless the model, quantization, context, runtime and workload are identical.
NVIDIA’s statement that workloads can reach up to 200 billion parameters describes the platform’s intended range. It does not establish that every model at that size will provide interactive latency, fit at full precision or support a chosen context window.
Troubleshoot the first setup systematically
The wizard cannot join the permanent network
- Confirm that you have completed the language, time zone and keyboard pages rather than closing the wizard early.
- For hotspot setup, remain connected to the SSID and password printed in the Quick Start Guide until the wizard confirms the permanent network.
- Allow the documented update and network stage up to 10 minutes before treating it as failed.
- If remote access is unavailable afterward, reconnect a display and keyboard or return to the local setup route.
A model will not load
- Check that the runtime build matches Spark’s ARM64 environment or, for Windows ARM64, the documented CUDA build.
- Confirm that the model and quantization fit the available unified memory with the selected context.
- Stop other inference servers and retry with one runtime.
- Use the runtime’s error output to distinguish a missing dependency, an incompatible binary and insufficient memory.
Windows reports missing CUDA DLLs
For the Windows ARM64 llama.cpp path, verify that you used the CUDA 13 build and that the CUDA 13.4 DLLs are in the same folder as llama-server.exe and llama-cli.exe. A correct model download does not compensate for DLLs stored elsewhere.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →A remote session stops working after an update
Use the local console if available, check DGX Dashboard and review the release notes for changes to out-of-box setup, over-the-air updates, NVIDIA Sync, NemoClaw onboarding or NCCL. These features change over time, so a procedure that worked on an earlier image may require a current release-specific adjustment.
Maintain a reproducible Spark development setup
Before a serious project, write down the Spark software release, operating-system image, runtime, model identifier, quantization, context setting and access method. Keep that record with the application or notebook. It makes a later update diagnosable and prevents a “works on the box” workflow from depending on undocumented GUI choices.
Re-check NVIDIA’s current DGX Spark release notes and developer recipes before copying commands into automation. The portal was updated on September 3, 2026, and NVIDIA notes that out-of-box setup, OTA behavior, NemoClaw onboarding, NVIDIA Sync and NCCL capabilities evolve with releases.
Quick Recap
A practical setup sequence
- Choose local display setup or temporary-hotspot setup.
- Connect power and network access, then complete language, time zone and keyboard settings.
- Wait for updates and network joining to finish, allowing up to 10 minutes.
- Open DGX Dashboard and confirm monitoring and JupyterLab access.
- Keep a local recovery method while testing SSH, NVIDIA Sync or remote desktop.
- Select one runtime: llama.cpp for control, Ollama for a simple command path, or LM Studio Bionic for graphical testing.
- Run one model, starting with the current Qwen3.8-27B recommendation for 24–128 GB systems when using NVIDIA’s llama.cpp recipe.
- Connect one application or notebook and record the working configuration.
- Add fine-tuning, image or video workflows, vLLM, SGLang or containers only after the baseline is stable.
- Attempt multi-Spark scaling only when one-device inference and recovery are predictable.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
