Flexible Chip Ports Could Reduce AI Networking Costs by 70-80%

Chips with flexible ports can adapt to various mesh dimensions, allowing for optimized connections within racks, to neighboring racks, and across superpods, unlike NVIDIA's fixed NVLink and PCIe categories. Utilizing fullmesh or hybrid (UB-Mesh) network topologies, as explored in the Ascend 950 paper, significantly reduces the need for switch chips and optical connectors, leading to substantial networking cost savings of 70-80%. This approach, while potentially harder to maintain, localizes high-bandwidth interactions, making it ideal for AI training workloads.

View Tweet→

Sources (1)

Semicon Twitter · Twitter · 8h ago

View tweet→

Credibility

Lead sourceSemicon Twitter

Corroboration1 distinct outlet

VerificationSingle source

Entities

NVIDIAmentioned

Unlock the full signal with Pro

Tessara synthesizes 80+ data sources into actionable intelligence for AI infrastructure investors.

Unlock with Pro→Sign in

← SignalsSignal

Watch7.0NetworkingSingle sourceactive

Updated 8h ago · 1 source

Flexible Chip Ports Could Reduce AI Networking Costs by 70-80%

Chips with flexible ports can adapt to various mesh dimensions, allowing for optimized connections within racks, to neighboring racks, and across superpods, unlike NVIDIA's fixed NVLink and PCIe categories. Utilizing fullmesh or hybrid (UB-Mesh) network topologies, as explored in the Ascend 950 paper, significantly reduces the need for switch chips and optical connectors, leading to substantial networking cost savings of 70-80%. This approach, while potentially harder to maintain, localizes high-bandwidth interactions, making it ideal for AI training workloads.

View Tweet→