Up to 75% of AI training runtime can be lost to networking rather than computation, according to research from Google and Microsoft. AMD's new Pensando Vulcano 800 AI NIC targets that bottleneck directly, promising faster job completion, greater fault tolerance and lower switching costs for large-scale AI clusters.
A Network Interface Card Built for GPU-Hungry Workloads
The Pensando Vulcano 800 AI NIC is designed to handle high-throughput AI training and distributed inference workloads. It slots into a standard rack and combines three 800Gbps NICs per GPU, delivering up to 2.4Tbps of total bandwidth to each GPU.
Since the network directly impacts compute utilisation, closing the gap between raw GPU power and actual usable throughput is central to getting value out of expensive AI infrastructure investments.
Balancing Scale-Out and Scale-Across Architecture
Raw bandwidth alone doesn't guarantee real-world cluster performance. To maximise efficiency, the Pensando NIC combines scale-out networking between nodes with scale-across networking between GPUs.
AMD Networking describes scale-out as the ability to distribute a workload across multiple nodes, expanding the effective GPU pool beyond what a single server can provide. For training, this means syncing workloads across hundreds of servers without slowdowns; for inference, it means handling heavy request traffic smoothly. Scale-across, meanwhile, extends performance beyond the rack to clusters that span multiple data centres, keeping bandwidth high and latency low over long distances.
"Think of a data centre rack the way you would think of an engine: every part has to work with every other part, or none of it performs the way it should."— Chitra Barvekar, Senior Silicon Design Engineer, AMD
Faster and More Resilient AI Clusters
AMD notes that throughput alone doesn't determine success under real-world conditions — job completion rates, fault tolerance and sustained daily use matter just as much. By eliminating retransmits, system stalls and recovery cycles, the Pensando NIC delivers up to 13% faster AI job completion times, based on AMD engineering silicon modelling and synthetic benchmark simulation.
The card also distributes traffic across independent data paths, resulting in up to 33% lower switching costs. If one link or component fails, system performance degrades gracefully rather than bringing an entire multi-million-dollar training job to a sudden halt. Paired with in-service diagnostics and rapid fault isolation, infrastructure teams can fix issues on the fly without scheduling disruptive, cluster-wide maintenance windows.