Adam Conway’s XDA test rendered a high-detail Blender scene on an HP OmniBook Ultra 16 with NVIDIA RTX Spark using 49.3 GB of scene memory. The render took about 12 minutes at 1080p and 64 samples. That result shows the tested system could handle this particular memory-heavy scene; it does not show that RTX Spark renders faster than an RTX 5090.
What the test measured
The system had a 20-core Grace CPU, a 6,144-core Blackwell GPU and 128 GB of shared LPDDR5X memory. XDA ran Blender 5.3 beta with the Cycles renderer and OptiX, NVIDIA’s ray-tracing software backend. Its GPU memory allocation was set to the default 64 GB, rather than the maximum 96 GB XDA says the NVIDIA app allows.
The test scene was a procedurally generated city. As its detail increased, Blender evaluated more geometry, and scene memory use rose sharply:
| Detail setting | Evaluated vertices | Peak scene memory | 1080p render time and samples |
| Level 4 | 6.65 million | 6.7 GB | 10–17 seconds at 48 samples |
| Level 6 | 106.3 million | 38.4 GB | 6.5 minutes at 16 samples |
| Level 6, with 11% of buildings at Level 7 | 139.3 million | 49.3 GB | About 12 minutes at 64 samples |
The 49.3 GB figure is the scene allocation. XDA reported 50.6 GB of total GPU memory use at peak, including about 1.3 GB already in use before rendering. The project file itself was under 2 MB; XDA attributed most of the memory demand to evaluated geometry and OptiX acceleration structures, which help speed up ray tracing.
Render times are tied to the settings in each row: both scene detail and sample count changed. At the highest detail, even a 16-sample render took more than 11 minutes. Conway said that geometry preparation and building acceleration structures accounted for much of the time, rather than sample tracing alone.
Why this is a capacity result, not a speed win
A desktop RTX 5090 has 32 GB of dedicated video memory, while the RTX 5090 Laptop GPU has 24 GB. Both figures are below the 49.3 GB scene allocation in XDA’s test. But that does not mean either GPU is categorically unable to render the scene: Cycles can move some allocations to system RAM. Other data must remain in GPU memory, and a render can fail if those remaining allocations do not fit.
Nor does the memory result establish faster rendering on RTX Spark. XDA reports 300 GB/s of memory bandwidth for the tested system, compared with 1,792 GB/s for the desktop RTX 5090, and says the RTX 5090 is faster for scenes that fit in its video memory. The test did not render the same city scene on an RTX 5090, so it offers no direct speed comparison between the two.
For general Blender use, Blender recommends 32 GB of Windows system RAM. NVIDIA Studio places an RTX 5060 Laptop GPU with 8 GB of video memory in its “Essential” category for medium-sized 3D scenes. Those are broad starting points, not guarantees for every project: the 49.3 GB scene was an unusually demanding workload, not a baseline requirement for Blender.
A separate Blender benchmark adds context
In a different test, DDay.it reported a score of 4,679 for RTX Spark in Blender 5.3 with OptiX on a 20-core, 128 GB Surface Laptop Ultra. Its cited RTX 5070 Laptop median was 4,005, and its RTX 5070 Ti Laptop median was 5,007. DDay noted that those comparison medians combine results from different Blender versions, operating systems and laptop power limits, so the figures are context rather than a controlled head-to-head comparison. They measure a separate benchmark, not the 49.3 GB city scene.
How a local model helped build the scene
Qwen3.8 Flash Next, running locally on the tested laptop, generated the city and the scripts used to create and render it. XDA reported that the process from prompt to the first full-size Cycles render took about 81 minutes without author intervention. The inference server was shut down before the final renders, so the reported scene render followed the model-assisted preparation rather than running alongside it.
NeoTeo’s earlier coverage of RTX Spark’s local AI workloads looks at a different use of the platform.
