My RTX Spark laptop rendered a Blender scene too big for an RTX 5090, and a local LLM on the same machine built it - XDA
A local LLM on an HP OmniBook Ultra 16 with RTX Spark generated a Blender scene requiring 49GB of GPU memory, exceeding the capacity of Nvidia's RTX 5090 and even the 48GB RTX PRO 5000 workstation card.
The HP OmniBook Ultra 16 with RTX Spark successfully rendered a complex Blender scene that required 49GB of GPU memory, exceeding the capacity of Nvidia's most powerful consumer graphics card. The scene, designed entirely by a local LLM running on the same machine, demonstrates the potential of unified memory architectures for creative workloads that would typically require expensive workstation GPUs.
The laptop features Nvidia's 20-core Grace CPU, a 6,144-core Blackwell GPU, and 128GB of LPDDR5X-9400 memory shared between components over a 256-bit bus with 300GB/s bandwidth. The RTX Spark reserves a portion of this memory for the GPU, which Nvidia calls a "carveout," and users can adjust it up to 96GB through the Nvidia app.
AI-designed city
Rather than designing the complex Blender scene himself, author Adam Conway used Qwen3.8 Flash Next, a local LLM with 125 billion parameters, to generate an abandoned futuristic city. The model, running via llama.cpp with a 262K context window, created a Python script using Blender's bpy API to generate a 15-by-15 grid of blocks containing around 13,600 objects including buildings, vehicles, vegetation, and lighting.
Everything you see in these renders, from the buildings to the materials and lighting, was generated by code written by a local LLM running on the same laptop.
The LLM even generated 18 camera positions automatically by analyzing the scene to find landmarks and framing appropriate shots. The entire process from initial prompt to first full render took approximately 81 minutes without human intervention.
Memory demands
The Blender scene featured configurable detail levels that significantly impacted memory usage. At the highest tested level, Blender evaluated 139.3 million vertices and used 49.3GB of GPU memory, bringing total reported usage to 50.6GB. Each 1080p frame at this detail level took around 12 minutes to render, with the laptop's fans running at full speed.
This memory requirement exceeds the 32GB of the RTX 5090 and even the 48GB RTX PRO 5000 workstation card. The 72GB version of the PRO 5000 costs $9,999, while the 96GB RTX PRO 6000 can cost twice that amount—more than the entire HP OmniBook Ultra 16 laptop itself.
Unified memory advantage
The RTX Spark's unified memory architecture allows it to handle workloads that would be impossible on conventional gaming laptops. While the dedicated GPU memory carveout benefits from optimizations, the system can overflow into shared memory when needed, though with some performance penalty compared to the dedicated region.
Blender support for RTX Spark's GPU rendering on Windows on Arm was added in September 2026, specifically built with the RTX Spark in mind and incorporating unified memory optimizations. This capability opens new possibilities for creative professionals who need powerful rendering capabilities in a portable form factor.
Related
