Industrial GPUs for Industrial Problems: Hands-On Access to the Latest NVIDIA GPUs in the Cloud

Industrial Problems Seminar

Tyler Whitehouse
Nvidia

Abstract

Modern GPUs aren’t just “faster graphics cards”. Architectural changes expand what’s possible by redefining the computations. For example, recent generations of NVIDIA Tensor Cores have added support for FP8 precision, structured-sparsity, and mix-precision calculations, thus contributing to larger models coupled with faster training and inference. These architectural changes bubble up to users as higher-level programming frameworks, e.g. pytorch, absorb the capabilities of the new version of CUDA.

Theoretically anyone can get up and running with the latest improvements in short order, especially if they have CUDA experience. The catch is access. You don’t find A100, H100, H200 or B200 GPUs on free tier or teaching cloud platforms. They’re also difficult to find on the larger cloud service providers because they’ve already been reserved by a group with the budget for yearly or multi-year contracts.

This is where NVIDIA’s Brev platform comes in. It’s a friendly GPU broker that provides access to a broad range of GPUs from a variety of large and small cloud providers. No waitlists or long-term contracts.

In this seminar, I’ll use Brev to spin up machines with different GPUs and run some basic benchmarks that highlight how architectural changes and configurations affect performance. All attendees will receive free credits to get started on Brev and experiment hands-on with the full range of NVIDIA GPUs and see the differences directly.