AetherScale • Neural VRAM Planner

Neural Rendering is the flashy experiment, but nobody wants to find out at 2 AM that the resolution they just queued doesn't fit. The Neural VRAM Planner is the pack's answer: a pure calculator that estimates how much free VRAM a Neural Rendering run at a given resolution will need, before you spend a second of GPU time. It has no GPU work and no runtime dependency - it's arithmetic plus a quick probe of the DLSS runtime state, wrapped in a node.

It's a small, honest tool, and that's its charm. You feed it a resolution and a couple of knobs; it hands back a JSON plan telling you the estimated surfaces in MB and, crucially, an estimated_required_free_mb number you can compare against your card's free VRAM.

The inputs that matter

• width / height - defaults to 3840×2160, which tells you exactly what the author was testing this against. Set them to your actual source frames.
• history_frames - how many frames of temporal history the Neural Rendering path keeps (default 2). More history, more surfaces.
• safety_margin_mb - the headroom buffer (default 2048). This is the one beginners actually tune: if the plan says you'd be cutting it close, raise the margin before you run, and lower it if you know the card has headroom to spare.
• measured_context_mb - the honest-tuning input. Leave it at 0 until you have a real measurement from an actual Neural Rendering run; once you do, the planner can fold reality into its estimate instead of guessing.

How the math works

Peek at the source and it's transparent: pixel count × a set of per-surface bytes-per-pixel assumptions for the visible buffers, scaled by history frames, plus your measured context and safety margin. Nothing hidden, no model-loading guesswork for the Diffusion side - this estimates the Neural Rendering workload's own surfaces. The output plan string also includes live dlssnr runtime info, so it doubles as a quick way to confirm the runtime state from within a workflow.

Installing and using it

Manager → search AetherScale, or clone into custom_nodes and restart, same as every node in this pack. Windows + RTX + Python 3.10+ if you want the downstream Neural Rendering to actually run, though this node alone is harmless on any platform.

The honest caveat

An estimate is an estimate. The planner's real value is the feedback loop: run it before a Neural Rendering job to sanity-check a resolution, then after the job read the actual numbers in the Neural Rendering stats output and feed the real context back into measured_context_mb. Do that a couple of times on your specific card and the planner becomes genuinely predictive for that card - until then, treat the output as a budget, not a prophecy. If you're just here for stable upscaling and restoration, you don't need this node at all; it exists to de-risk the experimental DLSS5 path.