Brief IA
›

NVIDIA: RTX Spark PCs Arrive with Up to 128GB

Creative AI·Tom Levy·

NVIDIA: RTX Spark PCs Arrive with Up to 128GB

NVIDIA: RTX Spark PCs Arrive with Up to 128GB
Key Takeaways
1ASUS, Dell, and MSI are marketing their compact RTX Spark computers running Windows.
2NVIDIA announces the local execution of language models with 120 billion parameters.
3Substance 3D Painter offers a native Arm version in public beta.
Why it matters : According to NVIDIA, compatible models executed locally keep processed files on the PC and avoid cloud API token fees for these tasks.

ASUS, Dell, and MSI are marketing their compact NVIDIA RTX Spark computers running Windows, equipped with up to 128 GB of unified memory. Aimed particularly at creators and developers, these machines support local models and AI agents. Their arrival is accompanied by updates in Adobe applications, Blender, and development tools.

Adobe and Blender Adapt Their Creative Tools

Substance 3D Painter 12.2 now features a native Arm version, offered in public beta and optimized for RTX Spark. Professionals in video game development, visual effects, or visualization can run it on compact machines as well as on laptops in the range, without a compatibility layer. Illustrator and InDesign also have their native versions for Windows on Arm. For Illustrator, NVIDIA claims performance gains of up to 30% compared to emulated execution.

In Premiere, TensorRT-RTX accelerates the Enhance Speech, Object Mask, Scene Edit Detection, and Auto Reframe functions. Effects and transitions from Film Impact extensions also benefit from the GPU. NVIDIA reports improved handling of Canon RAW, Red RAW, ARRI RAW, ARRICORE, ARRI HDE, as well as H.264 and HEVC in MXF, ProRes RAW, and long-GOP videos. The manufacturer announces over 50% faster processing for common deliverables in H.264.

Photoshop integrates TensorRT-RTX into its removal tool, while Camera Raw operates natively on RTX Spark: processing RAW files does not require emulation. During Adobe MAX, Adobe also showcased a new image processing engine that leverages the GPU more, particularly for interactive Gaussian blur and light leak filters. These developments are part of the initiative led by Adobe and NVIDIA to extend native operation on Windows on Arm and RTX acceleration.

On the Blender 5.3 side, DLSS 4.5 Ray Reconstruction joins the Cycles preview window. NVIDIA presents this feature as a way to examine lighting, reflections, and materials in real-time with quality close to the final result, rather than in a noisy preview. Its use requires this version of Blender and the latest Game Ready or Studio driver. The NVIDIA application offers driver availability alerts, a GPU control center to optimize applications and games, as well as access to the manufacturer's tools.

Machines Designed for Always-On Agents

The ASUS ProArt GR1X mini PC is among those computers designed, according to NVIDIA, to host personal agents 24/7. A dedicated machine can keep agents and scheduled tasks running when the laptop is taken away or occupied with another activity. Developers, creators, and AI enthusiasts are the target audiences for this combination of creative applications, development tools, image and video generation, and model tuning.

The configurations rely on a Grace processor with 20 cores and a Blackwell RTX GPU featuring 6,144 CUDA cores, accompanied by NVIDIA's AI software stack. The manufacturer claims a power of up to 1 petaflop for AI. The 10-gigabit Ethernet connection is intended for exchanging datasets, model files, and creative resources with network storage.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Among the capabilities claimed by NVIDIA are the local execution of language models with 120 billion parameters and real-time rendering of 3D scenes exceeding 90 GB. The manufacturer also announces real-time 4:2:2 video editing. The compact computers join the RTX Spark laptops already on sale, which are also targeted by the expansion of the software offering. Machines from Acer, GIGABYTE, and HP are expected to follow soon, along with the Microsoft Surface RTX Spark Dev Box.

Hermes Distributes Tasks Among Models

NVIDIA Switchyard is now integrated into Hermes: this tool automatically directs agent tasks to models running on the machine or in the cloud. According to NVIDIA, routine operations can be processed on the device, while more advanced reasoning needs may call on remote models to manage cloud-related expenses. The manufacturer specifies that local execution of compatible models keeps the relevant files and code on the PC and avoids cloud API billing for these processes.

Hermes Agent can, for instance, enable communication between Photoshop and ComfyUI, linking applications to local AI tools through commands expressed in natural language. NVIDIA describes the ability to convert sketches into images or animations, as well as to animate an image without leaving the device: selecting the subject, defining the movement, and then generating video. During Adobe MAX, participants used these processes to feed a collective digital landscape representing Miami.

The RTX Video Frame Generation and RTX Video HDR functions are also already accessible in ComfyUI. According to NVIDIA, they allow a clip generated at 60 frames per second to be converted into HDR for calibration, all within the same ComfyUI process on RTX Spark.

Development Tools and Microsoft Announcements

For developers on RTX Spark, CUDA is natively supported, and Linux tools are accessible via Windows Subsystem for Linux. Agent configuration can rely on Hermes Agent or NVIDIA NemoClaw models. Tuning compatible models can be done through Unsloth.

Microsoft also announced at Ignite that Windows ML can now run any model in GGUF format from Hugging Face using the llama.cpp engine. DxCGC, available as an experimental feature, allows game developers to compile and run machine learning models with DirectX 12, alongside graphics processing. Windows APIs add GPU support for AION Instruct and VSR, whose capabilities can be integrated into applications. A command-line interface is also available for building models for Windows ML.

For organizations, Microsoft announced at the same event an evolution of Windows Server aimed at bringing AI processing closer to enterprise applications and data. The publisher is extending compatibility with the latest AI chips for accelerated inference and the use of more capable models in internal environments. NVIDIA presents the new accelerator platforms as a multi-tiered offering in terms of price and capability for advanced development, local inference, and large-scale agents. The DGX Station for Windows, equipped with the GB300 Grace Blackwell Ultra Desktop superchip, is set to support Windows Server 2028.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.