{"id":92700,"date":"2025-05-19T13:07:00","date_gmt":"2025-05-19T04:07:00","guid":{"rendered":"https:\/\/kr-dev.rebellions.ai\/?p=92700"},"modified":"2025-08-24T11:58:36","modified_gmt":"2025-08-24T02:58:36","slug":"breaking-barriers-in-physical-ai-cosmos-runs-on-atom","status":"publish","type":"post","link":"https:\/\/dev-rbln-kr.locomotion.co.kr\/?p=92700","title":{"rendered":"Breaking Barriers in Physical AI: Cosmos Runs on ATOM\u2122"},"content":{"rendered":"\n
Multimodal AI is redefining the limits of infrastructure. Models are no longer confined to single input types\u2014they now span text, images, video, and more. At the frontier of this shift is Cosmos*: a diffusion based text-to-video generation model that integrates at least 5-6 distinct AI subsystems, including tokenizer, pre-trained world foundation models, and guardrail system (see Figure 1). Each subsystem brings its own architectural complexity and compute behavior. Traditionally, such workloads have been bound to high-end GPU platforms\u2014not because they are ideally suited, but because nothing else could handle the scale. GPUs remain powerful general-purpose accelerators, but their reliance on fixed software stacks and rigid memory hierarchies often make adaptation slow, inefficient, or overly manual. Cosmos is not just large\u2014it\u2019s structurally diverse and dynamically sequenced. Efficient execution isn\u2019t just about compute throughput. It\u2019s about system-level adaptability. And that\u2019s exactly what Rebellions set out to build.<\/p>\n\n\n
A Paradigm Shift: Enabling the Unexpected<\/strong><\/h2>\n\n\n\n
Rebellions is the first to run Cosmos-Predict1-7B in real time on a commercial NPU\u2014not by brute-forcing performance, but by building a stack that is inherently flexible and scalable. This wasn\u2019t a one-off integration.<\/p>\n\n\n\n
The Anatomy of Innovation: Full-Stack Mastery<\/strong><\/h2>\n\n\n\n
Cosmos ran on ATOM\u2122 because every layer of the stack\u2013from compiler to runtime to silicon\u2013was built to adapt to architectural diversity and scale with emerging models. Where others optimize for one model, we enable many. That adaptability isn\u2019t bolted on\u2014it\u2019s built in. Modern AI models demand infrastructure that evolves with them. GPU-based systems offer performance but often rely on legacy toolchains, fragmented software layers, and vendor-locked environments. Rebellions takes a system-level approach. By developing the full stack in-house\u2014from compiler to runtime to hardware\u2014we\u2019ve built a cohesive platform designed to absorb architectural diversity, not resist it. Instead of asking models to conform to infrastructure, we shape infrastructure to fit the model. This principle is what made Cosmos on ATOM\u2122 not just possible\u2014but replicable.<\/p>\n\n\n
\n[Figure 2. Rebellions Full-stack System Diagram (SW to HW)]<\/figcaption><\/figure>\n<\/div>\n\n\n
Software Innovations: Adaptability by Design<\/strong><\/h3>\n\n\n\n
\n
Modular compiler stack: Adaptability begins with modularity. Rebellions built a compiler stack designed for diversity: with a rich primitive library to support more than 200 distinct models, a front-end compiler that parses transformers, convolutions, and diffusion models, and a back-end that generates hardware-optimized code with minimal hand-tuning.<\/li>\n\n\n\n
Unified runtime orchestration: No two models behave alike\u2014and our runtime treats them accordingly. Rebellions\u2019 runtime dynamically orchestrates memory, compute, and kernel execution flows based on model behavior\u2014maximizing utilization while adapting to layer-wise variation in workloads like Cosmos. It also incorporates predictive DMA scheduling to preload data before it\u2019s needed\u2014ensuring that inputs arrive just-in time and compute stalls are avoided even under asynchronous, multi-stage workloads.<\/li>\n\n\n\n
Scalable communication infrastructure: Scalability isn\u2019t optional\u2014it\u2019s essential. Our \u201cRebellions Scalable Design\u201d enables multi-card inference through high-bandwidth communication layers and tensor-parallel execution. It\u2019s what allows Cosmos to span across NPUs without introducing latency or interconnect bottlenecks.<\/li>\n<\/ul>\n\n\n\n
Hardware Innovations: Purpose-Built for Flexibility<\/strong><\/h3>\n\n\n\n
\n
Flexible compute architecture: The hardware must move with the model. Rebellions engineered ATOM\u2122\u2019s compute cores to support a wide operator set and dynamic execution paths\u2014so that model-level flexibility translates seamlessly into silicon-level efficiency. This flexibility was essential for handling Cosmos\u2019 diverse operations without requiring model-specific kernel rewrites.<\/li>\n\n\n\n
Memory-compute coordination: Modern models don\u2019t compute linearly\u2014and neither should memory. Rebellions designed this coordination logic to track runtime behavior and adjust compute memory flow in real time\u2014making it inherently capable of handling bursty, irregular phases like those in Cosmos without special case engineering.<\/li>\n\n\n\n
High-bandwidth interconnect for multi-card scaling: Scaling across silicon shouldn\u2019t mean starting from scratch. Rebellions developed custom interconnect IP to enable high-throughput communication across NPUs. This ensures low-latency coordination for tensor-parallel workloads like Cosmos without introducing bottlenecks at scale.<\/li>\n<\/ul>\n\n\n\n
<\/div>\n\n\n\n<\/figure>\n\n\n\n
<\/div>\n\n\n
\n[Figure 3. From Prompt to Video: Cosmos-Predict1-7B Inference on ATOM\u2122]<\/figcaption><\/figure>\n<\/div>\n\n\n
![\"\"](\"https:\/\/kr-dev.rebellions.ai\/wp-content\/uploads\/2025\/05\/physical-ai_1_1.png\")
![\"\"](\"http:\/\/kr-dev.rebellions.ai\/wp-content\/uploads\/2025\/05\/3D-diagram_-runtime-module-v.3.1_color.png\")
![\"\"](\"http:\/\/kr-dev.rebellions.ai\/wp-content\/uploads\/2025\/05\/physical-ai_3@2x-1.png\")
<\/figure>\n\n\n\n<\/div>\n\n\n\n![\"\"](\"http:\/\/kr-dev.rebellions.ai\/wp-content\/uploads\/2025\/05\/pysical-ai_4@2x.png\")