AI

NVIDIA RTX Spark desktops: up to 128GB, from ASUS, Dell, MSI

Key takeaways

  • RTX Spark desktops pair a 20-core Grace CPU with a Blackwell GPU and up to 128GB of unified memory for 120-billion-parameter models locally
  • Windows on Arm support from Adobe and Blender 5.3 DLSS Ray Reconstruction matter as much as the hardware
  • NVIDIA gives no prices in this announcement, so cost per system and UK availability need checking with each vendor

NVIDIA's RTX Spark compact desktops are on sale now from ASUS, Dell and MSI, giving developers, creators and AI enthusiasts a dedicated Windows PC built to run large language models, image and video generation and always-on software agents locally. The systems pair a 20-core NVIDIA Grace CPU with a Blackwell RTX GPU that has 6,144 CUDA cores and up to 128GB of unified memory, delivering up to 1 petaflop of AI performance. The desktops join the RTX Spark laptops already on sale.

The pitch is a box that sits on a desk or a shelf, keeps working when a laptop is closed or travelling, and does not bill per token for the work it handles.

What it does well

The headline capability is memory. Unified memory means the CPU and GPU draw from one shared pool rather than the GPU being limited to its own separate video memory. NVIDIA says the hardware can run 120-billion-parameter language models locally and render 90GB+ 3D scenes in real time. Our explainer on why 128GB of shared memory matters more than your GPU goes into the mechanics.

Agents are the second selling point. Developers get native CUDA support and Linux tooling through Windows Subsystem for Linux, can set up agents with Hermes Agent or NVIDIA NemoClaw blueprints, and can fine-tune supported models using Unsloth. NVIDIA Switchyard, now available in Hermes, routes agent tasks automatically between local and cloud models. Routine jobs stay on the machine; harder reasoning goes to a cloud model. That split is a practical way to cap API spend without giving up access to frontier models.

The creative workflow demos are concrete. Hermes Agent can connect Photoshop and ComfyUI so that natural-language instructions turn a sketch into a rendered image or animation, or animate an image entirely on device. RTX Video Frame Generation and RTX Video HDR, already in ComfyUI, can then take a generated clip to 60 frames per second and convert it to high dynamic range for grading, all in one workflow.

Adobe has native Windows on Arm builds of Illustrator and InDesign, with Illustrator running up to 30% faster than its emulated version. Photoshop gains TensorRT-RTX acceleration for the Remove Tool and native Camera Raw. Premiere accelerates Enhance Speech, Object Mask, Scene Edit Detection and Auto Reframe, with the update said to give more than 50% quicker turnaround on everyday H.264 deliverables. Substance 3D Painter 12.2 runs natively in public beta. Blender 5.3 puts DLSS 4.5 Ray Reconstruction into the Cycles viewport, so artists see near-final-quality ray-traced previews instead of a noisy, slowly resolving image.

Where it falls short

Most of the performance figures here come from NVIDIA and Adobe, not independent testing. The "up to 1 petaflop" number is a peak AI rating, and the 128GB figure is an upper limit (the sources say 'up to 128GB'). The announcement does not say which memory options each vendor offers.

The Arm question is the bigger caveat. The platform depends on Windows on Arm, and the Adobe news is framed as an expanding ecosystem, with Substance Painter still in public beta and Adobe only previewing a next-generation image engine. Applications without native Arm builds run under emulation, which is exactly what the native releases are meant to avoid. Teams with a long tail of x86-only plug-ins or line-of-business software should test before committing.

Blender's new viewport feature also has requirements: Blender 5.3 plus the latest NVIDIA Game Ready or Studio Driver. And the cloud-cost savings only apply to supported models run locally; advanced reasoning still draws on cloud models.

How it compares

NVIDIA's own material points to its DGX Spark developer kit and the higher-end DGX Station as neighbours. The table below uses only what the announcement states.

SystemPositioningNotable detail
RTX Spark compact desktop (ASUS, Dell, MSI)Dedicated Windows PC for local AI, creative and dev work20-core Grace CPU, 6,144 CUDA cores, up to 128GB unified memory, 10-gigabit Ethernet
RTX Spark laptopsPortable version of the same platformAlready available; agents can stay on the desktop while the laptop travels
DGX Station for WindowsHigher-end platformGB300 Grace Blackwell Ultra Desktop Superchip; Windows Server 2028 support coming

For laptop buyers, our coverage of the Arm CPU, Blackwell GPU and 128GB laptop chip covers the mobile side of the family.

Pricing and value

The announcement gives no prices, in dollars or pounds, and no UK availability details. Buyers will need to check ASUS, Dell and MSI listings directly. The value argument NVIDIA makes is indirect: local processing avoids per-token cloud API charges for supported models, and locally processed code and files stay on the PC.

One cost worth planning for is networking. The 10-gigabit Ethernet port is there for moving large datasets, model files and creative assets to and from networked storage, so a desk running big models will want a switch and storage that can keep up.

Who should use it

The fit is strongest for developers building and testing agents who want WSL, CUDA and Unsloth fine-tuning on a Windows machine; for 3D and video studios working with large scenes or 4:2:2 footage; and for IT teams that want code and files kept on-device for supported models. Enterprise buyers should note the broader Windows push: Windows Server 2028 support is coming to DGX Station for Windows, and Windows ML now runs any GGUF model from Hugging Face through the llama.cpp backend, which widens the local-model options.

Those who should wait include anyone dependent on x86-only software, and anyone who needs the price and UK stock position before planning. Acer, GIGABYTE and HP mini PCs and the Microsoft Surface RTX Spark Dev Box are still to come, so buyers who prefer those vendors have a reason not to rush. Those weighing hardware against GPU memory limits should also check our VRAM tier list.

Verdict

Worth serious evaluation for developers and creative teams already leaning on local models or paying meaningful cloud API bills, particularly those whose tools now have native Arm builds. Hold off if the workflow relies on emulated x86 software, or until ASUS, Dell and MSI publish pricing and UK stock. Watch for independent benchmarks of the 120-billion-parameter and 90GB+ scene claims, and for the Acer, GIGABYTE, HP and Surface models that will show how far vendor competition moves the price.

Frequently asked questions

What is inside an NVIDIA RTX Spark desktop?

A 20-core NVIDIA Grace CPU, a Blackwell RTX GPU with 6,144 CUDA cores and up to 128GB of unified memory, rated at up to 1 petaflop of AI performance.

Which vendors sell RTX Spark desktops today?

ASUS, Dell and MSI. Acer, GIGABYTE and HP mini PCs are coming soon, along with the Microsoft Surface RTX Spark Dev Box.

More from Future Technology HQ