Training

Train the largest models on clusters designed for distributed workloads from the ground up. Training pairs frontier GPUs with a non-blocking InfiniBand fabric so your job scales close to linearly as you add nodes.

What you get

• Thousands of interconnected GPUs on a non-blocking InfiniBand fabric
• Checkpointing and automatic recovery from node failures
• Job scheduling with fair-share queues and priority tiers
• Throughput and utilisation metrics for every rank