Docs / Start Here / What this is
Start Here
What this is
A single 16 GB GPU running ComfyUI, shared with a research-paper RAG replica. What that buys you and what it costs.
The short version
One laptop-class GPU generates 2D game art on demand. You bring it up, generate, and hand the card back. Nothing is sent to a third-party service; there is no per-image charge; the weights are all open and stay on local disk.
The machine
| GPU | NVIDIA RTX 5080 Laptop, 16 GB VRAM, Blackwell (sm_120) |
| Power budget | ~90 W measured under load — roughly half a desktop 5080 |
| Host | 16-core CPU, 62 GB RAM, ~1.9 TB free, headless Linux |
| Stack | ComfyUI on Python 3.12 with PyTorch 2.11.0+cu130 |
That “Laptop” suffix is the single most important line in the table. A laptop 5080 is not a desktop 5080 with a smaller case — it is a different, power-limited part. Measured on this box, a warmed fp16 matrix multiply sustains 36.6 TFLOPS at 91 W with the SM clock sitting at 1290 MHz, against a desktop part’s ~2600 MHz boost.
In practice that is better than it sounds. Measured end to end at 1024² and 20 steps:
| Time | Peak VRAM | |
|---|---|---|
| SDXL + pixel-art LoRA | 8 s | 7.3 GB |
| FLUX.2 klein-4B (fp8) | 16 s | 8.6 GB |
Seconds per image, not fractions of a second — which is fine for batching sprites and tedious for turning a dial and waiting.
It is a shared card, not a dedicated one
The same GPU hosts two always-on services:
| Resident | VRAM | What it is |
|---|---|---|
| Papers-RAG replica query API | ~3.4 GB | A read-only copy of the lab’s research-paper corpus |
| Ollama holding a 4B LLM | ~3.9 GB | Pinned resident, answers questions against that corpus |
Together they hold ~7.3 GB of the 16 GB, leaving ~8.7 GB if you generate alongside them. That is enough for SDXL and comfortable for SD 1.5, but not enough for the larger models.
So the stack evicts them on demand. comfy-up stops both and hands you the full ~15.3 GB;
comfy-down puts them back. See Sharing the GPU
for why this is safe and
what specifically is not stopped.
What it is good at
- Sprites, item icons, portraits, and other small-canvas art in a consistent style
- Batch generation — dozens of variations while you do something else
- Tilesets and seamless textures
- Iterating on a look without a per-image bill
What it is not good at
- Interactive fiddling with the biggest models. At ~90 W, a large model at high resolution is a make-a-coffee operation, not a live dial-turning one.
- Being an art director. It generates candidates; a human still picks, fixes, and assembles.
- Perfect text. Legible words inside generated images remain unreliable.
- Editing. This box generates. Cleanup, palette work, and assembly happen on a desktop machine in Aseprite, Krita, or a tilemap editor — the GPU box is headless and has no GUI.
The honest cost
The electricity is negligible and the models are free. The real cost is your time: prompt iteration, culling bad generations, and hand-fixing the survivors. Treat the output as a fast draftsman, not a finished asset pipeline.
Source: content/start/what-this-is.md · maintained in the nuilab-aigaming repository.