Scalability of Memory subsystem

Presented by Isaiah Weiner (Weka IO)

Traditional AI and Generative AI are reliant upon storage system performance to reduce cycle times and maximize the effectiveness of downstream resources like GPUs. Join us as we explore an IO profiling of real AI customers. We will examine the impact of overlapping pipeline stages and the I/O Blender effect, and quantify the hidden costs of metadata. As we go, you will see how filesystem latency becomes the overall gate to pipeline progress, and how Generative AI micro-burst performance requirements defy averages. We will also investigate training epoch delays related to media silos, and share how WEKA solves many of these problems.