Finding the VRAM Sweet Spot: A Benchmarking Approach for Video Models in ComfyUI
MoreColors185 · reddit · 2026-08-02
A developer using an RTX 4070 Ti SUPER (16GB) explores VRAM utilization when running video generation models like Wan/SCAIL-2 in ComfyUI. The author breaks down VRAM consumption into a static baseline (model weights, encoders) and dynamic computation (resolution, frame count, attention, temporary tensors).
Core Concept & Testing Plan
- Formula Estimation: Proposes separating VRAM into baseline + dynamic usage, attempting to roughly estimate dynamic load using a formula like Pixel Load = width × height × frames.
- Benchmarking Methodology: Plans to conduct 6–9 controlled tests varying resolution, frame count, and sampling steps to record peak VRAM and runtime. The goal is to derive a practical model to find the sweet spot between quality, generation time, and VRAM limits.
The author also seeks community advice on the most accurate way to measure actual VRAM usage, comparing the reliability of nvidia-smi, ComfyUI’s built-in reporting, and PyTorch's allocated/reserved memory metrics.
Related event: Developer Proposes VRAM Calculation Formula for ComfyUI Video Generation(2 posts)→
More from Infra
- GitHub rewrote the 800k-line Copilot runtime in Rust using AI agents, cutting CPU and memory by over 90% — bibryam · 2026-10-16
- NVIDIA quietly delists the entry-level RTX 5050 from its marketplace, following the RTX 5090 — chemist_slime · 2026-10-16
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16