Scene is 1.4M tris, four area lights, one subsurface material. Cycles, 128 samples, 1080x1920. Local render on the 2050: out of memory at frame 1.
Instead of guessing, I measured.
Numbers
| Setting | Peak VRAM | Result |
|---|---|---|
| 128 samples, full res | 4.3 GB | OOM |
| 128 samples, 50% res | 2.1 GB | 6m 40s / frame |
| 64 samples + denoise, full | 4.1 GB | OOM |
| 64 samples, tiled 256px | 3.2 GB | 9m 10s / frame |
Tiling drops memory but the time goes up, which I didn’t expect until I realised the BVH gets rebuilt per tile when the scene doesn’t fit.
At nine minutes a frame, a 20-second clip at 24fps is three days of rendering. That’s not a workflow, that’s a hostage situation.
The actual conclusion
This isn’t a settings problem. The card is the constraint and no amount of tweaking moves it. Two options: cut scene complexity until it fits, or rent a GPU.
I priced the second one. An A5000 at roughly $0.30/hr does this scene in about 25 seconds a frame — the same 20-second clip is under an hour, so a few dollars. Cheaper than three days of not being able to use my own laptop.
Moving to cloud rendering tomorrow. Filed under broke because nothing shipped today, but
measuring properly was the point.