For years, AI services have been locked into expensive GPU cloud infrastructure, burdened by high costs, latency, and privacy risks. ZETIC.ai introduces a breakthrough: an end-to-end automated SDK that converts your existing AI models into fully optimized on-device apps within hours. By leveraging the NPUs already inside billions of smartphones, we enable companies to eliminate GPU servers entirely, cutting costs by up to 99% while delivering real-time, private, and scalable AI experiences. This session unveils how the future of AI is no longer in the cloud — it’s already in your pocket.

This demo explores how to achieve high-performance AI on Google's Tensor Processing Units (TPUs) using the JAX ecosystem, with a specific focus on image recognition workflows. We’ll begin with micro-benchmarks that showcase JAX's unique advantages for TPU-based computation, such as its single-program, multiple-data (SPMD) programming model, which is ideal for the TPU's systolic array architecture. This setup is designed for integration into larger, production-grade environments, such as those running on Kubernetes.

Rebellions introduces REBEL-Quad, the world’s first UCIe-Advanced AI accelerator designed for peta-scale inference. Built for efficiency at every layer, REBEL-Quad redefines the economics of AI data centers—delivering higher throughput and lower power consumption compared to GPU-based systems. This session will feature a demo showcase, highlighting how REBEL-Quad brings frontier-class performance into production, enabling enterprises to scale large-model workloads without the energy tax. Join us to see how next-generation chiplet architecture translates into real customer value: efficiency, scalability, and faster time-to-deployment.

Following the MLCommons MLPerf Inference v5.1 Results on the morning of Tuesday 9th September on the keynote stage, Miro Hodak, Senior Member of Technical Staff, AI Performance Engineering at AMD will deliver a detailed analysis of the results followed by a Q&A session from the audience. 

Distributed training jobs are brittle; a single node failure can halt progress and waste expensive GPU cycles. This technical demo dives into Cluster Director, focusing on how engineers can automate resilient, large-scale GPU infrastructure. We'll start with a declarative YAML configuration to define and provision a multi-node GPU cluster, optimized with the ideal network topology for NCCL communication. The core of the demo will be a live failure simulation. You will see Cluster Director automatically detect a preempted node, perform remediation, and maintain the integrity of the running workload with minimal disruption.

Large language models can now power capable software agents, yet real‑world success comes from disciplined engineering rather than flashy frameworks. Most reliable agents are built from simple, composable patterns instead of heavy abstractions.

The talk will introduce patterns to add complexity and autonomy only when it pays off. Attendees should leave with a practical decision framework for escalating from a single prompt to multi‑step agents, also keeping in mind guardrails for shipping trustworthy, cost‑effective agents at scale.