Each AI processor will have a PCIe interface connecting to one or more host CPUs through PCIe retimers and PCI switches. Modern AI models are data-hungry, computation-heavy beasts that need specialized hardware just to function, let alone perform at their best. That's the job of an AI server—a custom-built system that keeps AI applications fast, scalable, and efficient. Featuring 8 H100 GPUs and a remarkable 640 billion transistors, DGX servers deliver six times the AI performance of the previous generation, particularly excelling. AI accelerator servers are optimized to handle the processing required for different types of AI workloads, but what they have in common is the need to scale and connect multiple cards in a system and allow many processors to work together. Use cases include natural learning processing, predictive. Broadcom's Ethernet Adapters (also referred to as Ethernet NICs) along with Arista Networks' switches (based on Broadcom's DNX and XGS family of ASICs) leverage RDMA (Remote Direct Memory Access) to eliminate any connectivity bottlenecks and facilitate a high-throughput, low-latency transport. To support HPC workloads like AI/ML training, back-end networks deploy spine-leaf architecture where leaf switches connect to every spine switch.
