News

NVFP4 and FP8 Land Natively in ComfyUI: The New VRAM Math

4 September 2026 · RenderBob team

ComfyUI now supports NVFP4 and FP8 natively. The number format the models run in now determines how much VRAM a motion studio actually needs.

ComfyUI now supports NVFP4 and FP8 data formats natively. The number format the models run in now determines how much VRAM a motion studio actually needs.

On RTX 50-series cards, NVFP4 delivers roughly 2.5x the throughput while cutting VRAM use by about 60% versus the previous baseline. FP8 lands around 1.7x faster with roughly 40% less VRAM. NVIDIA also reports RTX performance in ComfyUI improved about 40% since September through driver and runtime optimisation. In a year of scarce, expensive VRAM, quantization is more headroom on the same card.

For motion work this matters most at the edges of feasibility. A video workflow that OOMs at 1080p in full precision may run comfortably in NVFP4, and a longer clip that was impossible on a 16GB card may suddenly fit. That is the difference between an artist waiting on a queue and an artist iterating in near real time.

NVFP4's biggest gains are tied to RTX 50-series hardware, so a mixed studio running older 30- and 40-series cards will see uneven benefit. Quantization can introduce quality trade-offs on some models and steps, so pin it in a tested, version-pinned workflow rather than flipping it on per artist by feel. The heaviest video models still push past consumer cards at high resolution and frame count, even quantized.

Which precision runs where is a scheduling decision. Put NVFP4 on your newest owned nodes, and send jobs that still overflow to cloud nodes. A control plane that knows each node's precision profile and VRAM ceiling can route a job to hardware that will complete it, instead of letting an artist discover the ceiling at 10pm. Quantization bought the studio headroom. How you spend that headroom across owned and cloud capacity is the remaining work.