GB200 NVL72 Topology-Aware Scheduling: How Slurm Unlocks Exascale Performance

As AI model scale and complexity continue to grow, fully leveraging the performance of modern accelerated computing infrastructure depends not only on the raw compute power of the hardware itself, but also on how workloads are placed and scheduled. NVIDIA's latest technical blog provides a detailed look at how Slurm topology-aware job scheduling can unlock Exascale-level computing performance on the GB200 NVL72 system.

GB200 NVL72: The Next-Generation Hyperscale AI Computing Platform

The NVIDIA GB200 NVL72 is a flagship computing platform designed for hyperscale AI training and inference workloads. The system integrates 72 Blackwell GPUs within a single rack, interconnected via NVLink to create a high-bandwidth, low-latency GPU communication domain. GPU-to-GPU data exchange speeds far exceed traditional network interconnect solutions, providing a solid hardware foundation for large-scale parallel computing.

Notably, the "G" in GB200 stands for Grace CPU, while the "B" stands for Blackwell GPU. NVL72 means that 72 GPUs are fully interconnected through NVLink Switch chips. Each Blackwell GPU features up to 192GB of HBM3e memory, giving the entire NVL72 system a total memory capacity exceeding 13TB. At FP8 precision, the system delivers AI compute performance of 1.44 ExaFLOPS — which is where the term "Exascale" comes from.