August 12, 2026 Kaleido: Algorithm-Hardware Co-Design for Video Diffusion Transformers Kaleido 通过挖掘视频潜在空间通道级时空相关性,提出通道级复用算法与可重构 PE 阵列加速器,最高 6.6× 加速、18.4× 能耗节省,PSNR 提升 >17 dB Video Diffusion Transformers Hardware Accelerator Algorithm-Hardware Co-Design Spatio-Temporal Correlation RoPE Encoding