Visko streams its Orbis live world model on Google Cloud TPUs in 30 days
Original titleHow Visko streams real-time world models on Google Cloud TPUs
AISummary
Visko worked with Google Cloud TPU engineers to move its Orbis live world model onto a single Cloud TPU slice in 30 days. The pipeline renders at 1.07x playback speed, using Ulysses context parallelism, Pallas splash attention, and a rewrite of the audio stage from torchax to pure JAX, which ran 5.8x faster.
AIWhy it matters
The post explains how a live video model stays under a real-time clock on TPUs, with concrete latency, parallelism, and framework choices a builder can reuse.
Source: Google Cloud · AI & Machine Learning · cloud.google.comPublished · added here