--- title: How Visko streams real-time world models on Google Cloud... description: AI startup Visko partnered with Google Cloud TPU engineers to port its Orbis live video world model pipeline onto a single Cloud TPU slice in 30 days, hitting... canonical: https://daily.dev/posts/how-visko-streams-real-time-world-models-on-google-cloud-tpus-uzztsxrap twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How Visko streams real-time world models on Google Cloud TPUs | daily.dev og:description: AI startup Visko partnered with Google Cloud TPU engineers to port its Orbis live video world model pipeline onto a single Cloud TPU slice in 30 days, hitting... og:url: https://daily.dev/posts/how-visko-streams-real-time-world-models-on-google-cloud-tpus-uzztsxrap og:image: https://api.daily.dev/og/posts/uzzTsXrAP.png og:image:alt: How Visko streams real-time world models on Google Cloud TPUs og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # How Visko streams real-time world models on Google Cloud TPUs **[Google Cloud](https://daily.dev/sources/gcp)** · 9 min read · 0 upvotes · 0 comments ## Summary AI startup Visko partnered with Google Cloud TPU engineers to port its Orbis live video world model pipeline onto a single Cloud TPU slice in 30 days, hitting 1.07x playback speed. The pipeline chains video diffusion, super-resolution upscaling, and audio generation in BF16 precision using JAX, torchax, and custom Pallas kernels. Key optimizations included Ulysses context parallelism across TPU chips (outperforming tensor parallelism and ring attention at batch size 1), a custom Pallas splash attention kernel from MaxDiffusion, and a 5.8x audio speedup achieved by rewriting the audio stage from torchax into native JAX. The team recommends using torchax as a fast bring-up bridge before selectively rewriting only latency-dominant components, keeping BF16 over FP8 to avoid quantization error accumulation across chunks, and caching compiled XLA graphs to speed worker restarts. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much faster did switching the audio model from torchax to native JAX make Visko's Orbis pipeline? Rewriting the audio generation stage from torchax to native JAX made it 5.8x faster, which was the change that brought the full multimodal pipeline under its real-time playback deadline. The video and super-resolution models stayed in torchax since they already met the latency target without a rewrite. _Teams weighing a full PyTorch-to-JAX rewrite versus a selective port can track real deployment tradeoffs like this on daily.dev._ ### Why is Ulysses context parallelism better than tensor parallelism for serving a single live video stream at batch size one? Ulysses context parallelism splits the token sequence across TPU chips instead of splitting model weights, using high-speed all-to-all interconnect exchanges so each chip handles the full sequence for a subset of attention heads. At batch size one, Visko found this approach faster than both tensor parallelism and ring attention for serving a single interactive live stream. _Engineers choosing a parallelism strategy for low-batch real-time inference can follow deployment patterns like this on daily.dev._ ### Why did Visko stick with BF16 precision instead of FP8 quantization for its live video world model? Native BF16 precision kept visual quality intact within the latency budget, while FP8 quantization was avoided because small quantization errors accumulate across chunks in a continuous streaming session, degrading output over time. This tradeoff mattered specifically because each chunk depends sequentially on the previous one's rolling memory of video latents. _Developers deciding between quantization levels for streaming AI workloads can compare real tradeoffs like this on daily.dev._ ## Similar posts on daily.dev - [How Lightricks trains video diffusion models at scale with JAX on TPU](https://daily.dev/posts/how-lightricks-trains-video-diffusion-models-at-scale-with-jax-on-tpu-cikcoqltz) · Google Cloud · 0 upvotes · 0 comments - [HeyGen x Google Cloud: Bringing Avatar IV to TPUs](https://daily.dev/posts/heygen-x-google-cloud-bringing-avatar-iv-to-tpus-qzakruxj5) · Google Developers · 0 upvotes · 0 comments - [Accelerating Spatio-Temporal Attention for Video Diffusion on TPUs](https://daily.dev/posts/accelerating-spatio-temporal-attention-for-video-diffusion-on-tpus-by8crmpum) · Google Developers · 0 upvotes · 0 comments --- Tags: [#genai](https://daily.dev/tags/genai), [#pytorch](https://daily.dev/tags/pytorch) [View this post on daily.dev](https://daily.dev/posts/how-visko-streams-real-time-world-models-on-google-cloud-tpus-uzztsxrap) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"How Visko streams real-time world models on Google Cloud TPUs","url":"https://daily.dev/posts/how-visko-streams-real-time-world-models-on-google-cloud-tpus-uzztsxrap","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-visko-streams-real-time-world-models-on-google-cloud-tpus-uzztsxrap"},"datePublished":"2026-10-15T16:03:03.708Z","dateModified":"2026-10-15T16:10:37.354Z","description":"AI startup Visko partnered with Google Cloud TPU engineers to port its Orbis live video world model pipeline onto a single Cloud TPU slice in 30 days, hitting...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/150f1a92c911be1a98914dd90afdbbe0?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/150f1a92c911be1a98914dd90afdbbe0?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Google Cloud","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Google Cloud","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/gcp","url":"https://daily.dev/sources/gcp"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/how-visko-streams-real-time-world-models-on-google-cloud-tpus-uzztsxrap","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"genai,pytorch","timeRequired":"PT9M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Google Cloud","item":"https://daily.dev/sources/gcp"},{"@type":"ListItem","position":3,"name":"How Visko streams real-time world models on Google Cloud TPUs"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/how-visko-streams-real-time-world-models-on-google-cloud-tpus-uzztsxrap#faq","mainEntity":[{"@type":"Question","name":"How much faster did switching the audio model from torchax to native JAX make Visko's Orbis pipeline?","acceptedAnswer":{"@type":"Answer","text":"Rewriting the audio generation stage from torchax to native JAX made it 5.8x faster, which was the change that brought the full multimodal pipeline under its real-time playback deadline. The video and super-resolution models stayed in torchax since they already met the latency target without a rewrite. Teams weighing a full PyTorch-to-JAX rewrite versus a selective port can track real deployment tradeoffs like this on daily.dev."}},{"@type":"Question","name":"Why is Ulysses context parallelism better than tensor parallelism for serving a single live video stream at batch size one?","acceptedAnswer":{"@type":"Answer","text":"Ulysses context parallelism splits the token sequence across TPU chips instead of splitting model weights, using high-speed all-to-all interconnect exchanges so each chip handles the full sequence for a subset of attention heads. At batch size one, Visko found this approach faster than both tensor parallelism and ring attention for serving a single interactive live stream. Engineers choosing a parallelism strategy for low-batch real-time inference can follow deployment patterns like this on daily.dev."}},{"@type":"Question","name":"Why did Visko stick with BF16 precision instead of FP8 quantization for its live video world model?","acceptedAnswer":{"@type":"Answer","text":"Native BF16 precision kept visual quality intact within the latency budget, while FP8 quantization was avoided because small quantization errors accumulate across chunks in a continuous streaming session, degrading output over time. This tradeoff mattered specifically because each chunk depends sequentially on the previous one's rolling memory of video latents. Developers deciding between quantization levels for streaming AI workloads can compare real tradeoffs like this on daily.dev."}}]} ```