Generative climate emulation on CPU with consistency models

Generative climate model emulators offer an efficient alternative to computationally expensive Earth System Models (ESMs) for exploring future scenario uncertainty in impacts, adaptation, and mitigation studies. However, without GPU acceleration, their iterative sampling procedures remain a computational bottleneck for more modest but widely available hardware. Here, we show that a consistency model can emulate impact-relevant climate variables on a CPU at a rate of 6 samples per second, providing a $\times$280 speedup over a baseline diffusion model. The model reproduces internal variability and forced trends under a warming scenario unseen during training with limited loss in distributional fidelity. These results provide important evidence that consistency models can make generative climate emulation practical without dedicated GPU acceleration.