CodiLime’s Post

The CPU does exactly three things in our AI firewall. Then it walks away and never comes back. Phase 1, runs once: allocate the shared buffers, capture the GPU's CUDA graph, launch the firewall app on the DPU. Setup done. Phase 2, runs forever: the DPU and GPU handle every packet between themselves. No CPU dispatch loop. The GPU replays its own captured graph and schedules its own work. Most "AI in infrastructure" means wiring a model in at the application layer, where everyone has the same tools. Designing a data path the CPU can leave entirely is a different altitude of engineering, and far fewer teams can build there. That's the layer this one lives on. Want to know how to get there? Read this: https://lnkd.in/dPCeZh2s