49.3GB of GPU memory was used to render a procedurally generated Blender city on Nvidia's RTX Spark laptop, pushing total reported usage to 50.6GB and completing 1080p frames in about 12 minutes.
128GB of unified LPDDR5X memory let the system exceed its 64GB default GPU carveout without a PCIe fallback, giving Blender more usable memory than an RTX 5090's 32GB VRAM and even the 48GB RTX Pro 5000.
Blender 5.3 beta made the test possible by adding Windows-on-Arm CUDA and OptiX support for RTX Spark; the scene itself scaled to 139.3 million vertices at the highest detail level.
Qwen 3.8 Flash Next built the 13,600-object city locally in about 81 minutes, showing the same laptop could handle both a large LLM workflow and the final GPU render.
The result does not make RTX Spark faster than an RTX 5090, but it shows how large unified-memory laptops can run workstation-class scenes that would otherwise be constrained by VRAM.