Elevate networking for the age of Gen AI

Before

AI server networking component sprawl

• NICs, PCIe and Rail Switches connected as stovepipes
• Limited bandwidth and fault-tolerance

Congestion-prone data movement across GPUs

• Multiple device hops
• Unpredictable load distribution, incast-prone

Ballooning AI Cluster TCO

• Stranded resources with lower effective bandwidth
• GPU link failure stalls entire job

A New Approach

Collective 8X NIC for collective GPUs

• 8X scale-out RDMA bandwidth
• Efficient traffic distribution across GPUs

Cut down GPU data movement latencies

• Up to 66% fewer device hops in large GPU clusters
• Congestion-free network-to-GPU traffic

Boost AI cluster operational efficiency

• Eliminates job failures due to link flaps
• 50% lower network power, 80% fewer components

Deploy easily

Connect seamlessly within AI clusters

• Using 100/200/400/800GbE optical or copper links
• To spine switches with high radix multipathing

Scale quickly without changing AI software

• Plug and play with standard ibverbs and xCCL
• Transparent scale-out networking acceleration

Experience it

• Flexible mix of CPU, GPU and CXL.Mem/SSD nodes
• Familiar interface for AI and HPC frameworks
• SuperNIC and Memory expansion SDKs
• Ideal for fast piloting and deployment