A Software Tweak Just Made AI Training 1.7x Faster

Researchers have found a way to speed up training of Joint Embedding Predictive Architectures (JEPAs) — which the authors describe as becoming a core representation-learning primitive and a building block for latent world models across vision, video, audio, brain dynamics, and time series. Current JEPA training pipelines are inefficient: each input is executed through multiple mask-specific branches, with redundant target-side work and memory-bound token routing. Their method, M-JEPA, is a mask-aware execution architecture that restructures JEPA training without changing the learning objective, separating mask-independent computation from mask-dependent routing. It enables shared context encoder execution, fused token routing and slicing with backward support, sparse target encoder execution over the union of target tokens, and masked patch embedding for sparse inputs. Implemented for five JEPA variants and evaluated on NVIDIA A100 GPUs, M-JEPA achieves up to 1.7x end-to-end training speedup for 2-10 masks. Separately, with masked patch embedding, it achieves 4.75x patch-embedding speedup at high sparsity. The authors say these results show that execution restructuring, rather than changes to the JEPA objective, is a key lever for efficient JEPA training.