Case Study: Nvidia achieves near-linear GPU scaling for medical AI with GigaIO FabreX

Nvidia - Customer Case Study

Nvidia needed a more flexible, high-performance way to scale GPU-based medical AI training and inference workloads without sacrificing efficiency. To address this challenge, they worked with GigaIO and its FabreX PCIe composable infrastructure to connect servers and GPU resources through a rack-scale PCIe fabric.

GigaIO implemented a composed architecture using FabreX network adapters, a TOR switch, and two accelerator pooling appliances with 16 NVIDIA GPUs. The solution used GPUDirect RDMA and PCIe-native communication to achieve near-linear scaling, with measured efficiency reaching 84% at 16 GPUs while supporting fast, software-defined reconfiguration and improved resource utilization.

Nvidia