This presentation shares the high-stakes story of Meta's transformation from lagging behind the industry in GPU platform time-to-production (TTP) to becoming an industry leader. We'll begin by explaining the significance of TTP, providing a brief history of our New Product Introductions (NPIs), including AI platforms. A crash course on NPIs will follow, setting the stage for a deeper dive into the key lessons learned from these experiences. Next, we'll walk you through the major changes that drove our dramatic improvement in TTP and conclude with an overview of our current challenges and future work.

The growth of data and need for increased power efficiency are leading to innovative storage solutions. HDDs have been growing in density, but not performance, and TLC flash remains at a price point that is restrictive for scaling. QLC technology addresses these challenges by forming a middle tier between HDDs and TLC SSDs. At Meta we are deploying high density QLC racks at scale, driving innovation to overcome interesting software and hardware challenges, and promoting ecosystem alignment in this evolving storage space.

In this talk, we will discuss the challenges of running ultra-low latency Large Language Model (LLM) inference at scale. We will cover the unique challenges of LLM inference, such as large model sizes, KV Caching. We will also discuss the challenges of scaling LLM inference to handle large volumes of requests, including the need for hardware, efficient scale up, and new routing architectures. Finally, we will present some of our recent work on addressing these challenges, including our development of inference infrastructure at Union.

The rapid advancement of AI has necessitated a fundamental shift in infrastructure, moving from homogenous workloads that fit within a single server to multi-host workloads requiring tight container coordination across multiple servers. This talk explores the motivations and design principles behind this shift, focusing on the implementation of first-class support for gang scheduling at all layers of the system. We delve into the key components of this design, including Twine and the Resource Allowance System (RAS), and examine how they enable AI serving schemes that employ various forms of parallelism—such as pipeline, context, tensor, and expert parallelism—requiring container shared fate properties and network topology-aware allocation. By addressing these challenges, we aim to provide insights into building scalable and reliable systems that meet the demands of modern AI workloads.

Both public clouds and our hyperscale private cloud have evolved into complex infrastructures with millions of servers spanning numerous global data center regions. Leaving users to manage the complexity of deploying global services across regions incurs significant operational toil and often leads to suboptimal outcomes. Users must select regions, align global traffic routing with service deployments and ensure disaster preparedness, while optimizing cost. They must continually repeat these processes to adapt to workload changes. To eliminate these manual burdens, we introduce the Global Service Placer (GSP), which autonomously places services across regions based on user-defined latency SLOs, abstracting away details of regions from users, such as their geo-distribution and resource availability, while optimizing efficiency. The generality and efficacy of GSP has been demonstrated via onboarding a diverse set of big and complex services. We provide a case study for highly complex AI inference workload and show significant GPU savings.

Meta's recommendation systems rely on "freshness" – the speed at which user interaction signals are ingested, trained, and utilized. To improve model freshness, Meta developed solutions addressing scaling, serving footprints, and diverse architectures.

We outline Nvidia's experience managing a large-scale internal GPU compute platform spanning multiple heterogeneous clusters. The platform supports thousands of users and hundreds of project accounts, handling a diverse mix of training and batch inference workloads across various research fields. We focus on three key challenges: researcher productivity, resource utilization, and operational efficiency. To improve researcher productivity, we emphasize fair scheduling and workload resilience. To keep resource utilization high, we discuss strategies to maintain high occupancy. On the operational efficiency front, we highlight our scheduler simulation capabilities, which enable safe testing of changes without affecting production workloads. The presentation concludes with key lessons learned and our vision for future improvements.