Researchers introduce Phantom temporal caching accelerator

Researchers introduce Phantom, an algorithm-architecture co-designed accelerator that uses temporal caching to speed up diffusion model inference. By treating temporal similarity as differential sparsity, Phantom recomputes only significant outlier tokens between adjacent timesteps while reusing the rest from a cache. Evaluations show the accelerator achieves an average 4.0x speedup and 5.5x energy saving over standard GPUs and prior accelerators while preserving generation quality. Open article →