DriveNets

Large language model inference is increasingly becoming a systems problem. Many techniques exist, but three performance drivers matter most: - Shrink it: Reduce the KV-cache footprint - Move it: Transfer KV state efficiently - Reuse it: Avoid unnecessary recomputation One of the most important parts of infrastructure optimization is Key-Value (KV) cache management. This white paper examines how KV-cache size, movement, and reuse affect inference performance. As models grow, context windows expand, and inference architectures become more distributed, delivering better performance is no longer simply a matter of deploying faster GPUs. The infrastructure surrounding those GPUs increasingly determines how efficiently they can serve tokens.

The next phase of AI will reward those that optimize their entire AI cluster. The most expensive idle asset in AI today is a GPU waiting on data, and no amount of compute budget can fix a cluster bottlenecked by data starvation. This blog explores how cross-vendor full-stack optimization, as demonstrated by the recent DriveNets, AMD and VAST Data collaboration, provides the most complete blueprint for large-scale AI inference infrastructure.

DriveNets and AMD published a validated Reference Architecture that maximize the performance of AI clusters built with AMD Instinct MI355X GPUs, AMD Pollara NICs, and DriveNets scale-out and frontend solutions. The integrated solution from AMD and DriveNets eliminates networking bottlenecks by performing end-to-end networking optimizations across the entire AI stack, including collective communication libraries, transport protocols, NICs, the network fabric, and system-level orchestration. It maximizes GPU utilization and accelerates TTFT, optimizing AI cluster economics. The joint reference architecture enables: ✔️ 15% faster Time to First Token (TTFT) at the single-node level ✔️ 12–16% faster TTFT at scale ✔️ Up to 5% higher multi-node throughput ✔️ Stronger performance as concurrency increases

DriveNets will be at SC26 in Chicago! Learn more about how our AI Ethernet fabric supports InfiniBand-level performance, and about our work with AMD. DriveNets' AI networking solutions deliver: Highest E2E performance for AMD based on full-stack optimization DriveNets' Open, lossless Ethernet fabric delivers approximately 5% higher throughput and 10-15% lower TTFT for MI355X GPU clusters, compared to publicly available industry results Fastest deployment time without complex fine-tuning and automated orchestration

Tata Communications Lanka is carrying live production traffic over a network powered by DriveNets Network Cloud. The deployment in Sri Lanka puts the DriveNets NCPlite white box router into production for Tata Communications Lanka Internet POP network. DriveNets' Technical Delivery Team delivered the solution end to end – from on-site training through design, implementation, HLD/LLD, migration, testing and go-live. The deployment demonstrates how DriveNets can support demanding, loss-sensitive services.

DriveNets and AMD published a validated Reference Architecture that maximize the performance of AI clusters built with AMD Instinct MI355X GPUs, AMD Pollara NICs, and DriveNets scale-out and frontend solutions. The integrated solution from AMD and DriveNets eliminates networking bottlenecks by performing end-to-end networking optimizations across the entire AI stack, including collective communication libraries, transport protocols, NICs, the network fabric, and system-level orchestration. The joint reference architecture enables: ✔️ 15% faster Time to First Token (TTFT) at the single-node level ✔️ 12–16% faster TTFT at scale ✔️ Up to 5% higher multi-node throughput ✔️ Stronger performance as concurrency increases