¿Por qué WanGP usa más CPU en tu RTX 3090?

You recently noticed that WanGP seems to be utilizing significantly more CPU power compared to older versions. You were looking at your system metrics on an RTX 3090 and found yourself wondering if the software itself had become more CPU-intensive.

The answer is nuanced. It isn’t simply that WanGP has become ‘more CPU-dependent’ in its core calculations, but rather that the architecture of the software has shifted dramatically. The 2026 updates have introduced a host of new optimizations—specifically regarding VRAM and RAM management, quantization profiles, and offloading techniques—that naturally shift more work onto your processor.

The Role of Memory Management

Consider the context of your hardware: an RTX 3090 with its 24 GB of VRAM. When you utilize specific WanGP optimization profiles to keep models like MiniMax H3 or LTX-2.3 within that 24 GB limit, the software is forced to manage data more aggressively.

This creates a workflow where the system constantly cycles between the GPU and your main RAM: GPU → CPU RAM → GPU. While the actual inference happens on your RTX 3090, the CPU ends up handling a considerable amount of peripheral work. It is responsible for managing memory copies, preparing tensors, handling VAE processing, and managing video data.

New Features Add Overhead

Beyond just memory management, WanGP has incorporated features that inherently increase the workload on your CPU. Recent versions include more intensive video processing, Media Flow capabilities, previews, and sliding window management. Furthermore, there is a clear correlation between video length and RAM requirements.

A Proposed Experiment

To move past speculation, you can run a concrete experiment using your current setup. You should compare the exact same generation of MiniMax H3 across both your older WanGP version and the new one. By monitoring CPU percentage, total RAM usage, VRAM consumption, GPU power draw, and inference time per token, you will be able to pinpoint exactly what is driving that spike in activity.

Whether it’s a genuine increase in CPU workloads or simply the overhead of these new memory profiles becomes clear once you run those side-by-side comparisons.