When training models on ultra-large datasets, one of the biggest challenges is low GPU utilization. These powerful processors are often underutilized due to inefficient I/O and data access. This mismatch between computation and storage leads to wasted GPU resources, low performance, and high cloud storage costs. The rise of generative AI and GPU scarcity is only making this problem worse.
In this webinar, Tarik and Beinan discuss strategies for transforming idle GPUs into optimal powerhouses. They will focus on cost-effective management of ultra-large datasets for AI and analytics.
What you will learn:
- The challenges of I/O stalls leading to low GPU utilization for model training
- High-performance, high-throughput data access (I/O) strategies
- The benefits of using an on-demand data access layer over your storage
- How Uber addresses managing ultra-large datasets using high-density storage and caching
When training models on ultra-large datasets, one of the biggest challenges is low GPU utilization. These powerful processors are often underutilized due to inefficient I/O and data access. This mismatch between computation and storage leads to wasted GPU resources, low performance, and high cloud storage costs. The rise of generative AI and GPU scarcity is only making this problem worse.
In this webinar, Tarik and Beinan discuss strategies for transforming idle GPUs into optimal powerhouses. They will focus on cost-effective management of ultra-large datasets for AI and analytics.
What you will learn:
- The challenges of I/O stalls leading to low GPU utilization for model training
- High-performance, high-throughput data access (I/O) strategies
- The benefits of using an on-demand data access layer over your storage
- How Uber addresses managing ultra-large datasets using high-density storage and caching
Video:
Presentation slides:
Complete the form below to access the full overview:
.png)
Videos
TorchTitan is a proof-of-concept for Large-scale LLM training using native PyTorch. It is a repo that showcases PyTorch's latest distributed training features in a clean, minimal codebase.
In this talk, Tianyu will share TorchTitan’s design and optimizations for the Llama 3.1 family of LLMs, spanning 8 billion to 405 billion parameters, and showcase its performance, composability, and scalability.
As large-scale machine learning becomes increasingly GPU-centric, modern high-performance hardware like NVMe storage and RDMA networks (InfiniBand or specialized NICs) are becoming more widespread. To fully leverage these resources, it’s crucial to build a balanced architecture that avoids GPU underutilization. In this talk, we will explore various strategies to address this challenge by effectively utilizing these advanced hardware components. Specifically, we will present experimental results from building a Kubernetes-native distributed caching layer, utilizing NVMe storage and high-speed RDMA networks to optimize data access for PyTorch training.