A kernel-centric path to real-time video generation on Trainium
Amazon Science
Reactor collaborated with Amazon Neuron Science to optimize real-time interactive video generation models on AWS Trainium using a kernel-centric workflow. It enabled a 3D RoPE kernel to drop from 5 seconds to 1.8 milliseconds and improved Rolling Forcing to sustain 16 frames per second for streaming generation. As a result, Trainium can run long-sequence autoregressive diffusion video workloads with lower latency by handling dynamic shapes, cache operations, and attention efficiently at the kernel level.
Why it matters
Using the Neuron Kernel Interface, a Reactor–AWS collaboration tackled the dynamic shapes, memory access patterns, and cache management that make real-time autoregressive diffusion hard—building techniques that generalize across models.
